bevy/crates/bevy_ecs/src/lib.rs

1745 lines
54 KiB
Rust
Raw Normal View History

#![warn(missing_docs)]
#![doc = include_str!("../README.md")]
#[cfg(target_pointer_width = "16")]
compile_error!("bevy_ecs cannot safely compile for a 16-bit platform.");
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
pub mod archetype;
pub mod bundle;
pub mod change_detection;
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
pub mod component;
pub mod entity;
pub mod event;
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
pub mod query;
#[cfg(feature = "bevy_reflect")]
pub mod reflect;
pub mod removal_detection;
pub mod schedule;
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
pub mod storage;
pub mod system;
pub mod world;
2020-07-17 02:20:51 +00:00
use bevy_utils::HashMap for better performance. TypeId is predefined … (#7642) …u64, so hash safety is not a concern # Objective - While reading the code, just noticed the BundleInfo's HashMap is std::collections::HashMap, which uses a slow but safe hasher. ## Solution - Use bevy_utils::HashMap instead benchmark diff (I run several times in a linux box, the perf improvement is consistent, though numbers varies from time to time, I paste my last run result here): ``` bash cargo bench -- spawn Compiling bevy_ecs v0.9.0 (/home/lishuo/developer/pr/bevy/crates/bevy_ecs) Compiling bevy_app v0.9.0 (/home/lishuo/developer/pr/bevy/crates/bevy_app) Compiling benches v0.1.0 (/home/lishuo/developer/pr/bevy/benches) Finished bench [optimized] target(s) in 1m 17s Running benches/bevy_ecs/change_detection.rs (/home/lishuo/developer/pr/bevy/benches/target/release/deps/change_detection-86c5445d0dc34529) Gnuplot not found, using plotters backend Running benches/bevy_ecs/benches.rs (/home/lishuo/developer/pr/bevy/benches/target/release/deps/ecs-e49b3abe80bfd8c0) Gnuplot not found, using plotters backend spawn_commands/2000_entities time: [153.94 µs 159.19 µs 164.37 µs] change: [-14.706% -11.050% -6.9633%] (p = 0.00 < 0.05) Performance has improved. spawn_commands/4000_entities time: [328.77 µs 339.11 µs 349.11 µs] change: [-7.6331% -3.9932% +0.0487%] (p = 0.06 > 0.05) No change in performance detected. spawn_commands/6000_entities time: [445.01 µs 461.29 µs 477.36 µs] change: [-16.639% -13.358% -10.006%] (p = 0.00 < 0.05) Performance has improved. spawn_commands/8000_entities time: [657.94 µs 677.71 µs 696.95 µs] change: [-8.8708% -5.2591% -1.6847%] (p = 0.01 < 0.05) Performance has improved. get_or_spawn/individual time: [452.02 µs 466.70 µs 482.07 µs] change: [-17.218% -14.041% -10.728%] (p = 0.00 < 0.05) Performance has improved. get_or_spawn/batched time: [291.12 µs 301.12 µs 311.31 µs] change: [-12.281% -8.9163% -5.3660%] (p = 0.00 < 0.05) Performance has improved. spawn_world/1_entities time: [81.668 ns 84.284 ns 86.860 ns] change: [-12.251% -6.7872% -1.5402%] (p = 0.02 < 0.05) Performance has improved. spawn_world/10_entities time: [789.78 ns 821.96 ns 851.95 ns] change: [-19.738% -14.186% -8.0733%] (p = 0.00 < 0.05) Performance has improved. spawn_world/100_entities time: [7.9906 µs 8.2449 µs 8.5013 µs] change: [-12.417% -6.6837% -0.8766%] (p = 0.02 < 0.05) Change within noise threshold. spawn_world/1000_entities time: [81.602 µs 84.161 µs 86.833 µs] change: [-13.656% -8.6520% -3.0491%] (p = 0.00 < 0.05) Performance has improved. Found 1 outliers among 100 measurements (1.00%) 1 (1.00%) high mild Benchmarking spawn_world/10000_entities: Warming up for 500.00 ms Warning: Unable to complete 100 samples in 4.0s. You may wish to increase target time to 4.0s, enable flat sampling, or reduce sample count to 70. spawn_world/10000_entities time: [813.02 µs 839.76 µs 865.41 µs] change: [-12.133% -6.1970% -0.2302%] (p = 0.05 < 0.05) Change within noise threshold. ``` --- ## Changelog > This section is optional. If this was a trivial fix, or has no externally-visible impact, you can delete this section. - use bevy_utils::HashMap for Bundles::bundle_ids ## Migration Guide > This section is optional. If there are no breaking changes, you can delete this section. - Not a breaking change, hashmap is internal impl.
2023-02-15 04:19:26 +00:00
use std::any::TypeId;
pub use bevy_ptr as ptr;
/// Most commonly used re-exported types.
2020-07-17 02:20:51 +00:00
pub mod prelude {
#[doc(hidden)]
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
#[cfg(feature = "bevy_reflect")]
pub use crate::reflect::{AppTypeRegistry, ReflectComponent, ReflectResource};
#[doc(hidden)]
2020-07-17 02:20:51 +00:00
pub use crate::{
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
bundle::Bundle,
change_detection::{DetectChanges, DetectChangesMut, Mut, Ref},
component::Component,
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
entity::Entity,
Migrate engine to Schedule v3 (#7267) Huge thanks to @maniwani, @devil-ira, @hymm, @cart, @superdump and @jakobhellermann for the help with this PR. # Objective - Followup #6587. - Minimal integration for the Stageless Scheduling RFC: https://github.com/bevyengine/rfcs/pull/45 ## Solution - [x] Remove old scheduling module - [x] Migrate new methods to no longer use extension methods - [x] Fix compiler errors - [x] Fix benchmarks - [x] Fix examples - [x] Fix docs - [x] Fix tests ## Changelog ### Added - a large number of methods on `App` to work with schedules ergonomically - the `CoreSchedule` enum - `App::add_extract_system` via the `RenderingAppExtension` trait extension method - the private `prepare_view_uniforms` system now has a public system set for scheduling purposes, called `ViewSet::PrepareUniforms` ### Removed - stages, and all code that mentions stages - states have been dramatically simplified, and no longer use a stack - `RunCriteriaLabel` - `AsSystemLabel` trait - `on_hierarchy_reports_enabled` run criteria (now just uses an ad hoc resource checking run condition) - systems in `RenderSet/Stage::Extract` no longer warn when they do not read data from the main world - `RunCriteriaLabel` - `transform_propagate_system_set`: this was a nonstandard pattern that didn't actually provide enough control. The systems are already `pub`: the docs have been updated to ensure that the third-party usage is clear. ### Changed - `System::default_labels` is now `System::default_system_sets`. - `App::add_default_labels` is now `App::add_default_sets` - `CoreStage` and `StartupStage` enums are now `CoreSet` and `StartupSet` - `App::add_system_set` was renamed to `App::add_systems` - The `StartupSchedule` label is now defined as part of the `CoreSchedules` enum - `.label(SystemLabel)` is now referred to as `.in_set(SystemSet)` - `SystemLabel` trait was replaced by `SystemSet` - `SystemTypeIdLabel<T>` was replaced by `SystemSetType<T>` - The `ReportHierarchyIssue` resource now has a public constructor (`new`), and implements `PartialEq` - Fixed time steps now use a schedule (`CoreSchedule::FixedTimeStep`) rather than a run criteria. - Adding rendering extraction systems now panics rather than silently failing if no subapp with the `RenderApp` label is found. - the `calculate_bounds` system, with the `CalculateBounds` label, is now in `CoreSet::Update`, rather than in `CoreSet::PostUpdate` before commands are applied. - `SceneSpawnerSystem` now runs under `CoreSet::Update`, rather than `CoreStage::PreUpdate.at_end()`. - `bevy_pbr::add_clusters` is no longer an exclusive system - the top level `bevy_ecs::schedule` module was replaced with `bevy_ecs::scheduling` - `tick_global_task_pools_on_main_thread` is no longer run as an exclusive system. Instead, it has been replaced by `tick_global_task_pools`, which uses a `NonSend` resource to force running on the main thread. ## Migration Guide - Calls to `.label(MyLabel)` should be replaced with `.in_set(MySet)` - Stages have been removed. Replace these with system sets, and then add command flushes using the `apply_system_buffers` exclusive system where needed. - The `CoreStage`, `StartupStage, `RenderStage` and `AssetStage` enums have been replaced with `CoreSet`, `StartupSet, `RenderSet` and `AssetSet`. The same scheduling guarantees have been preserved. - Systems are no longer added to `CoreSet::Update` by default. Add systems manually if this behavior is needed, although you should consider adding your game logic systems to `CoreSchedule::FixedTimestep` instead for more reliable framerate-independent behavior. - Similarly, startup systems are no longer part of `StartupSet::Startup` by default. In most cases, this won't matter to you. - For example, `add_system_to_stage(CoreStage::PostUpdate, my_system)` should be replaced with - `add_system(my_system.in_set(CoreSet::PostUpdate)` - When testing systems or otherwise running them in a headless fashion, simply construct and run a schedule using `Schedule::new()` and `World::run_schedule` rather than constructing stages - Run criteria have been renamed to run conditions. These can now be combined with each other and with states. - Looping run criteria and state stacks have been removed. Use an exclusive system that runs a schedule if you need this level of control over system control flow. - For app-level control flow over which schedules get run when (such as for rollback networking), create your own schedule and insert it under the `CoreSchedule::Outer` label. - Fixed timesteps are now evaluated in a schedule, rather than controlled via run criteria. The `run_fixed_timestep` system runs this schedule between `CoreSet::First` and `CoreSet::PreUpdate` by default. - Command flush points introduced by `AssetStage` have been removed. If you were relying on these, add them back manually. - Adding extract systems is now typically done directly on the main app. Make sure the `RenderingAppExtension` trait is in scope, then call `app.add_extract_system(my_system)`. - the `calculate_bounds` system, with the `CalculateBounds` label, is now in `CoreSet::Update`, rather than in `CoreSet::PostUpdate` before commands are applied. You may need to order your movement systems to occur before this system in order to avoid system order ambiguities in culling behavior. - the `RenderLabel` `AppLabel` was renamed to `RenderApp` for clarity - `App::add_state` now takes 0 arguments: the starting state is set based on the `Default` impl. - Instead of creating `SystemSet` containers for systems that run in stages, simply use `.on_enter::<State::Variant>()` or its `on_exit` or `on_update` siblings. - `SystemLabel` derives should be replaced with `SystemSet`. You will also need to add the `Debug`, `PartialEq`, `Eq`, and `Hash` traits to satisfy the new trait bounds. - `with_run_criteria` has been renamed to `run_if`. Run criteria have been renamed to run conditions for clarity, and should now simply return a bool. - States have been dramatically simplified: there is no longer a "state stack". To queue a transition to the next state, call `NextState::set` ## TODO - [x] remove dead methods on App and World - [x] add `App::add_system_to_schedule` and `App::add_systems_to_schedule` - [x] avoid adding the default system set at inappropriate times - [x] remove any accidental cycles in the default plugins schedule - [x] migrate benchmarks - [x] expose explicit labels for the built-in command flush points - [x] migrate engine code - [x] remove all mentions of stages from the docs - [x] verify docs for States - [x] fix uses of exclusive systems that use .end / .at_start / .before_commands - [x] migrate RenderStage and AssetStage - [x] migrate examples - [x] ensure that transform propagation is exported in a sufficiently public way (the systems are already pub) - [x] ensure that on_enter schedules are run at least once before the main app - [x] re-enable opt-in to execution order ambiguities - [x] revert change to `update_bounds` to ensure it runs in `PostUpdate` - [x] test all examples - [x] unbreak directional lights - [x] unbreak shadows (see 3d_scene, 3d_shape, lighting, transparaency_3d examples) - [x] game menu example shows loading screen and menu simultaneously - [x] display settings menu is a blank screen - [x] `without_winit` example panics - [x] ensure all tests pass - [x] SubApp doc test fails - [x] runs_spawn_local tasks fails - [x] [Fix panic_when_hierachy_cycle test hanging](https://github.com/alice-i-cecile/bevy/pull/120) ## Points of Difficulty and Controversy **Reviewers, please give feedback on these and look closely** 1. Default sets, from the RFC, have been removed. These added a tremendous amount of implicit complexity and result in hard to debug scheduling errors. They're going to be tackled in the form of "base sets" by @cart in a followup. 2. The outer schedule controls which schedule is run when `App::update` is called. 3. I implemented `Label for `Box<dyn Label>` for our label types. This enables us to store schedule labels in concrete form, and then later run them. I ran into the same set of problems when working with one-shot systems. We've previously investigated this pattern in depth, and it does not appear to lead to extra indirection with nested boxes. 4. `SubApp::update` simply runs the default schedule once. This sucks, but this whole API is incomplete and this was the minimal changeset. 5. `time_system` and `tick_global_task_pools_on_main_thread` no longer use exclusive systems to attempt to force scheduling order 6. Implemetnation strategy for fixed timesteps 7. `AssetStage` was migrated to `AssetSet` without reintroducing command flush points. These did not appear to be used, and it's nice to remove these bottlenecks. 8. Migration of `bevy_render/lib.rs` and pipelined rendering. The logic here is unusually tricky, as we have complex scheduling requirements. ## Future Work (ideally before 0.10) - Rename schedule_v3 module to schedule or scheduling - Add a derive macro to states, and likely a `EnumIter` trait of some form - Figure out what exactly to do with the "systems added should basically work by default" problem - Improve ergonomics for working with fixed timesteps and states - Polish FixedTime API to match Time - Rebase and merge #7415 - Resolve all internal ambiguities (blocked on better tools, especially #7442) - Add "base sets" to replace the removed default sets.
2023-02-06 02:04:50 +00:00
event::{Event, EventReader, EventWriter, Events},
query::{Added, AnyOf, Changed, Has, Or, QueryState, With, Without},
removal_detection::RemovedComponents,
schedule::{
apply_deferred, apply_state_transition, common_conditions::*, Condition,
IntoSystemConfigs, IntoSystemSet, IntoSystemSetConfigs, NextState, OnEnter, OnExit,
OnTransition, Schedule, Schedules, State, StateTransitionEvent, States, SystemSet,
},
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
system::{
Commands, Deferred, In, IntoSystem, Local, NonSend, NonSendMut, ParallelCommands,
ParamSet, Query, ReadOnlySystem, Res, ResMut, Resource, System, SystemParamFunction,
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
},
Allow disjoint mutable world access via `EntityMut` (#9419) # Objective Fix #4278 Fix #5504 Fix #9422 Provide safe ways to borrow an entire entity, while allowing disjoint mutable access. `EntityRef` and `EntityMut` are not suitable for this, since they provide access to the entire world -- they are just helper types for working with `&World`/`&mut World`. This has potential uses for reflection and serialization ## Solution Remove `EntityRef::world`, which allows it to soundly be used within queries. `EntityMut` no longer supports structural world mutations, which allows multiple instances of it to exist for different entities at once. Structural world mutations are performed using the new type `EntityWorldMut`. ```rust fn disjoint_system( q2: Query<&mut A>, q1: Query<EntityMut, Without<A>>, ) { ... } let [entity1, entity2] = world.many_entities_mut([id1, id2]); *entity1.get_mut::<T>().unwrap() = *entity2.get().unwrap(); for entity in world.iter_entities_mut() { ... } ``` --- ## Changelog - Removed `EntityRef::world`, to fix a soundness issue with queries. + Removed the ability to structurally mutate the world using `EntityMut`, which allows it to be used in queries. + Added `EntityWorldMut`, which is used to perform structural mutations that are no longer allowed using `EntityMut`. ## Migration Guide **Note for maintainers: ensure that the guide for #9604 is updated accordingly.** Removed the method `EntityRef::world`, to fix a soundness issue with queries. If you need access to `&World` while using an `EntityRef`, consider passing the world as a separate parameter. `EntityMut` can no longer perform 'structural' world mutations, such as adding or removing components, or despawning the entity. Additionally, `EntityMut::world`, `EntityMut::world_mut` , and `EntityMut::world_scope` have been removed. Instead, use the newly-added type `EntityWorldMut`, which is a helper type for working with `&mut World`. --------- Co-authored-by: Alice Cecile <alice.i.cecile@gmail.com>
2023-08-28 23:30:59 +00:00
world::{EntityMut, EntityRef, EntityWorldMut, FromWorld, World},
2020-07-17 02:20:51 +00:00
};
}
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
pub use bevy_utils::all_tuples;
Modular Rendering (#2831) This changes how render logic is composed to make it much more modular. Previously, all extraction logic was centralized for a given "type" of rendered thing. For example, we extracted meshes into a vector of ExtractedMesh, which contained the mesh and material asset handles, the transform, etc. We looked up bindings for "drawn things" using their index in the `Vec<ExtractedMesh>`. This worked fine for built in rendering, but made it hard to reuse logic for "custom" rendering. It also prevented us from reusing things like "extracted transforms" across contexts. To make rendering more modular, I made a number of changes: * Entities now drive rendering: * We extract "render components" from "app components" and store them _on_ entities. No more centralized uber lists! We now have true "ECS-driven rendering" * To make this perform well, I implemented #2673 in upstream Bevy for fast batch insertions into specific entities. This was merged into the `pipelined-rendering` branch here: #2815 * Reworked the `Draw` abstraction: * Generic `PhaseItems`: each draw phase can define its own type of "rendered thing", which can define its own "sort key" * Ported the 2d, 3d, and shadow phases to the new PhaseItem impl (currently Transparent2d, Transparent3d, and Shadow PhaseItems) * `Draw` trait and and `DrawFunctions` are now generic on PhaseItem * Modular / Ergonomic `DrawFunctions` via `RenderCommands` * RenderCommand is a trait that runs an ECS query and produces one or more RenderPass calls. Types implementing this trait can be composed to create a final DrawFunction. For example the DrawPbr DrawFunction is created from the following DrawCommand tuple. Const generics are used to set specific bind group locations: ```rust pub type DrawPbr = ( SetPbrPipeline, SetMeshViewBindGroup<0>, SetStandardMaterialBindGroup<1>, SetTransformBindGroup<2>, DrawMesh, ); ``` * The new `custom_shader_pipelined` example illustrates how the commands above can be reused to create a custom draw function: ```rust type DrawCustom = ( SetCustomMaterialPipeline, SetMeshViewBindGroup<0>, SetTransformBindGroup<2>, DrawMesh, ); ``` * ExtractComponentPlugin and UniformComponentPlugin: * Simple, standardized ways to easily extract individual components and write them to GPU buffers * Ported PBR and Sprite rendering to the new primitives above. * Removed staging buffer from UniformVec in favor of direct Queue usage * Makes UniformVec much easier to use and more ergonomic. Completely removes the need for custom render graph nodes in these contexts (see the PbrNode and view Node removals and the much simpler call patterns in the relevant Prepare systems). * Added a many_cubes_pipelined example to benchmark baseline 3d rendering performance and ensure there were no major regressions during this port. Avoiding regressions was challenging given that the old approach of extracting into centralized vectors is basically the "optimal" approach. However thanks to a various ECS optimizations and render logic rephrasing, we pretty much break even on this benchmark! * Lifetimeless SystemParams: this will be a bit divisive, but as we continue to embrace "trait driven systems" (ex: ExtractComponentPlugin, UniformComponentPlugin, DrawCommand), the ergonomics of `(Query<'static, 'static, (&'static A, &'static B, &'static)>, Res<'static, C>)` were getting very hard to bear. As a compromise, I added "static type aliases" for the relevant SystemParams. The previous example can now be expressed like this: `(SQuery<(Read<A>, Read<B>)>, SRes<C>)`. If anyone has better ideas / conflicting opinions, please let me know! * RunSystem trait: a way to define Systems via a trait with a SystemParam associated type. This is used to implement the various plugins mentioned above. I also added SystemParamItem and QueryItem type aliases to make "trait stye" ecs interactions nicer on the eyes (and fingers). * RenderAsset retrying: ensures that render assets are only created when they are "ready" and allows us to create bind groups directly inside render assets (which significantly simplified the StandardMaterial code). I think ultimately we should swap this out on "asset dependency" events to wait for dependencies to load, but this will require significant asset system changes. * Updated some built in shaders to account for missing MeshUniform fields
2021-09-23 06:16:11 +00:00
/// A specialized hashmap type with Key of [`TypeId`]
use bevy_utils::HashMap for better performance. TypeId is predefined … (#7642) …u64, so hash safety is not a concern # Objective - While reading the code, just noticed the BundleInfo's HashMap is std::collections::HashMap, which uses a slow but safe hasher. ## Solution - Use bevy_utils::HashMap instead benchmark diff (I run several times in a linux box, the perf improvement is consistent, though numbers varies from time to time, I paste my last run result here): ``` bash cargo bench -- spawn Compiling bevy_ecs v0.9.0 (/home/lishuo/developer/pr/bevy/crates/bevy_ecs) Compiling bevy_app v0.9.0 (/home/lishuo/developer/pr/bevy/crates/bevy_app) Compiling benches v0.1.0 (/home/lishuo/developer/pr/bevy/benches) Finished bench [optimized] target(s) in 1m 17s Running benches/bevy_ecs/change_detection.rs (/home/lishuo/developer/pr/bevy/benches/target/release/deps/change_detection-86c5445d0dc34529) Gnuplot not found, using plotters backend Running benches/bevy_ecs/benches.rs (/home/lishuo/developer/pr/bevy/benches/target/release/deps/ecs-e49b3abe80bfd8c0) Gnuplot not found, using plotters backend spawn_commands/2000_entities time: [153.94 µs 159.19 µs 164.37 µs] change: [-14.706% -11.050% -6.9633%] (p = 0.00 < 0.05) Performance has improved. spawn_commands/4000_entities time: [328.77 µs 339.11 µs 349.11 µs] change: [-7.6331% -3.9932% +0.0487%] (p = 0.06 > 0.05) No change in performance detected. spawn_commands/6000_entities time: [445.01 µs 461.29 µs 477.36 µs] change: [-16.639% -13.358% -10.006%] (p = 0.00 < 0.05) Performance has improved. spawn_commands/8000_entities time: [657.94 µs 677.71 µs 696.95 µs] change: [-8.8708% -5.2591% -1.6847%] (p = 0.01 < 0.05) Performance has improved. get_or_spawn/individual time: [452.02 µs 466.70 µs 482.07 µs] change: [-17.218% -14.041% -10.728%] (p = 0.00 < 0.05) Performance has improved. get_or_spawn/batched time: [291.12 µs 301.12 µs 311.31 µs] change: [-12.281% -8.9163% -5.3660%] (p = 0.00 < 0.05) Performance has improved. spawn_world/1_entities time: [81.668 ns 84.284 ns 86.860 ns] change: [-12.251% -6.7872% -1.5402%] (p = 0.02 < 0.05) Performance has improved. spawn_world/10_entities time: [789.78 ns 821.96 ns 851.95 ns] change: [-19.738% -14.186% -8.0733%] (p = 0.00 < 0.05) Performance has improved. spawn_world/100_entities time: [7.9906 µs 8.2449 µs 8.5013 µs] change: [-12.417% -6.6837% -0.8766%] (p = 0.02 < 0.05) Change within noise threshold. spawn_world/1000_entities time: [81.602 µs 84.161 µs 86.833 µs] change: [-13.656% -8.6520% -3.0491%] (p = 0.00 < 0.05) Performance has improved. Found 1 outliers among 100 measurements (1.00%) 1 (1.00%) high mild Benchmarking spawn_world/10000_entities: Warming up for 500.00 ms Warning: Unable to complete 100 samples in 4.0s. You may wish to increase target time to 4.0s, enable flat sampling, or reduce sample count to 70. spawn_world/10000_entities time: [813.02 µs 839.76 µs 865.41 µs] change: [-12.133% -6.1970% -0.2302%] (p = 0.05 < 0.05) Change within noise threshold. ``` --- ## Changelog > This section is optional. If this was a trivial fix, or has no externally-visible impact, you can delete this section. - use bevy_utils::HashMap for Bundles::bundle_ids ## Migration Guide > This section is optional. If there are no breaking changes, you can delete this section. - Not a breaking change, hashmap is internal impl.
2023-02-15 04:19:26 +00:00
type TypeIdMap<V> = rustc_hash::FxHashMap<TypeId, V>;
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
#[cfg(test)]
mod tests {
use crate as bevy_ecs;
Use lifetimed, type erased pointers in bevy_ecs (#3001) # Objective `bevy_ecs` has large amounts of unsafe code which is hard to get right and makes it difficult to audit for soundness. ## Solution Introduce lifetimed, type-erased pointers: `Ptr<'a>` `PtrMut<'a>` `OwningPtr<'a>'` and `ThinSlicePtr<'a, T>` which are newtypes around a raw pointer with a lifetime and conceptually representing strong invariants about the pointee and validity of the pointer. The process of converting bevy_ecs to use these has already caught multiple cases of unsound behavior. ## Changelog TL;DR for release notes: `bevy_ecs` now uses lifetimed, type-erased pointers internally, significantly improving safety and legibility without sacrificing performance. This should have approximately no end user impact, unless you were meddling with the (unfortunately public) internals of `bevy_ecs`. - `Fetch`, `FilterFetch` and `ReadOnlyFetch` trait no longer have a `'state` lifetime - this was unneeded - `ReadOnly/Fetch` associated types on `WorldQuery` are now on a new `WorldQueryGats<'world>` trait - was required to work around lack of Generic Associated Types (we wish to express `type Fetch<'a>: Fetch<'a>`) - `derive(WorldQuery)` no longer requires `'w` lifetime on struct - this was unneeded, and improves the end user experience - `EntityMut::get_unchecked_mut` returns `&'_ mut T` not `&'w mut T` - allows easier use of unsafe API with less footguns, and can be worked around via lifetime transmutery as a user - `Bundle::from_components` now takes a `ctx` parameter to pass to the `FnMut` closure - required because closure return types can't borrow from captures - `Fetch::init` takes `&'world World`, `Fetch::set_archetype` takes `&'world Archetype` and `&'world Tables`, `Fetch::set_table` takes `&'world Table` - allows types implementing `Fetch` to store borrows into world - `WorldQuery` trait now has a `shrink` fn to shorten the lifetime in `Fetch::<'a>::Item` - this works around lack of subtyping of assoc types, rust doesnt allow you to turn `<T as Fetch<'static>>::Item'` into `<T as Fetch<'a>>::Item'` - `QueryCombinationsIter` requires this - Most types implementing `Fetch` now have a lifetime `'w` - allows the fetches to store borrows of world data instead of using raw pointers ## Migration guide - `EntityMut::get_unchecked_mut` returns a more restricted lifetime, there is no general way to migrate this as it depends on your code - `Bundle::from_components` implementations must pass the `ctx` arg to `func` - `Bundle::from_components` callers have to use a fn arg instead of closure captures for borrowing from world - Remove lifetime args on `derive(WorldQuery)` structs as it is nonsensical - `<Q as WorldQuery>::ReadOnly/Fetch` should be changed to either `RO/QueryFetch<'world>` or `<Q as WorldQueryGats<'world>>::ReadOnly/Fetch` - `<F as Fetch<'w, 's>>` should be changed to `<F as Fetch<'w>>` - Change the fn sigs of `Fetch::init/set_archetype/set_table` to match respective trait fn sigs - Implement the required `fn shrink` on any `WorldQuery` implementations - Move assoc types `Fetch` and `ReadOnlyFetch` on `WorldQuery` impls to `WorldQueryGats` impls - Pass an appropriate `'world` lifetime to whatever fetch struct you are for some reason using ### Type inference regression in some cases rustc may give spurrious errors when attempting to infer the `F` parameter on a query/querystate this can be fixed by manually specifying the type, i.e. `QueryState::new::<_, ()>(world)`. The error is rather confusing: ```rust= error[E0271]: type mismatch resolving `<() as Fetch<'_>>::Item == bool` --> crates/bevy_pbr/src/render/light.rs:1413:30 | 1413 | main_view_query: QueryState::new(world), | ^^^^^^^^^^^^^^^ expected `bool`, found `()` | = note: required because of the requirements on the impl of `for<'x> FilterFetch<'x>` for `<() as WorldQueryGats<'x>>::Fetch` note: required by a bound in `bevy_ecs::query::QueryState::<Q, F>::new` --> crates/bevy_ecs/src/query/state.rs:49:32 | 49 | for<'x> QueryFetch<'x, F>: FilterFetch<'x>, | ^^^^^^^^^^^^^^^ required by this bound in `bevy_ecs::query::QueryState::<Q, F>::new` ``` --- Made with help from @BoxyUwU and @alice-i-cecile Co-authored-by: Boxy <supbscripter@gmail.com>
2022-04-27 23:44:06 +00:00
use crate::prelude::Or;
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
use crate::{
bundle::Bundle,
change_detection::Ref,
component::{Component, ComponentId},
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
entity::Entity,
Rename `WorldQueryData` & `WorldQueryFilter` to `QueryData` & `QueryFilter` (#10779) # Rename `WorldQueryData` & `WorldQueryFilter` to `QueryData` & `QueryFilter` Fixes #10776 ## Solution Traits `WorldQueryData` & `WorldQueryFilter` were renamed to `QueryData` and `QueryFilter`, respectively. Related Trait types were also renamed. --- ## Changelog - Trait `WorldQueryData` has been renamed to `QueryData`. Derive macro's `QueryData` attribute `world_query_data` has been renamed to `query_data`. - Trait `WorldQueryFilter` has been renamed to `QueryFilter`. Derive macro's `QueryFilter` attribute `world_query_filter` has been renamed to `query_filter`. - Trait's `ExtractComponent` type `Query` has been renamed to `Data`. - Trait's `GetBatchData` types `Query` & `QueryFilter` has been renamed to `Data` & `Filter`, respectively. - Trait's `ExtractInstance` type `Query` has been renamed to `Data`. - Trait's `ViewNode` type `ViewQuery` has been renamed to `ViewData`. - Trait's `RenderCommand` types `ViewWorldQuery` & `ItemWorldQuery` has been renamed to `ViewData` & `ItemData`, respectively. ## Migration Guide Note: if merged before 0.13 is released, this should instead modify the migration guide of #10776 with the updated names. - Rename `WorldQueryData` & `WorldQueryFilter` trait usages to `QueryData` & `QueryFilter` and their respective derive macro attributes `world_query_data` & `world_query_filter` to `query_data` & `query_filter`. - Rename the following trait type usages: - Trait's `ExtractComponent` type `Query` to `Data`. - Trait's `GetBatchData` type `Query` to `Data`. - Trait's `ExtractInstance` type `Query` to `Data`. - Trait's `ViewNode` type `ViewQuery` to `ViewData`' - Trait's `RenderCommand` types `ViewWolrdQuery` & `ItemWorldQuery` to `ViewData` & `ItemData`, respectively. ```rust // Before #[derive(WorldQueryData)] #[world_query_data(derive(Debug))] struct EmptyQuery { empty: (), } // After #[derive(QueryData)] #[query_data(derive(Debug))] struct EmptyQuery { empty: (), } // Before #[derive(WorldQueryFilter)] struct CustomQueryFilter<T: Component, P: Component> { _c: With<ComponentC>, _d: With<ComponentD>, _or: Or<(Added<ComponentC>, Changed<ComponentD>, Without<ComponentZ>)>, _generic_tuple: (With<T>, With<P>), } // After #[derive(QueryFilter)] struct CustomQueryFilter<T: Component, P: Component> { _c: With<ComponentC>, _d: With<ComponentD>, _or: Or<(Added<ComponentC>, Changed<ComponentD>, Without<ComponentZ>)>, _generic_tuple: (With<T>, With<P>), } // Before impl ExtractComponent for ContrastAdaptiveSharpeningSettings { type Query = &'static Self; type Filter = With<Camera>; type Out = (DenoiseCAS, CASUniform); fn extract_component(item: QueryItem<Self::Query>) -> Option<Self::Out> { //... } } // After impl ExtractComponent for ContrastAdaptiveSharpeningSettings { type Data = &'static Self; type Filter = With<Camera>; type Out = (DenoiseCAS, CASUniform); fn extract_component(item: QueryItem<Self::Data>) -> Option<Self::Out> { //... } } // Before impl GetBatchData for MeshPipeline { type Param = SRes<RenderMeshInstances>; type Query = Entity; type QueryFilter = With<Mesh3d>; type CompareData = (MaterialBindGroupId, AssetId<Mesh>); type BufferData = MeshUniform; fn get_batch_data( mesh_instances: &SystemParamItem<Self::Param>, entity: &QueryItem<Self::Query>, ) -> (Self::BufferData, Option<Self::CompareData>) { // .... } } // After impl GetBatchData for MeshPipeline { type Param = SRes<RenderMeshInstances>; type Data = Entity; type Filter = With<Mesh3d>; type CompareData = (MaterialBindGroupId, AssetId<Mesh>); type BufferData = MeshUniform; fn get_batch_data( mesh_instances: &SystemParamItem<Self::Param>, entity: &QueryItem<Self::Data>, ) -> (Self::BufferData, Option<Self::CompareData>) { // .... } } // Before impl<A> ExtractInstance for AssetId<A> where A: Asset, { type Query = Read<Handle<A>>; type Filter = (); fn extract(item: QueryItem<'_, Self::Query>) -> Option<Self> { Some(item.id()) } } // After impl<A> ExtractInstance for AssetId<A> where A: Asset, { type Data = Read<Handle<A>>; type Filter = (); fn extract(item: QueryItem<'_, Self::Data>) -> Option<Self> { Some(item.id()) } } // Before impl ViewNode for PostProcessNode { type ViewQuery = ( &'static ViewTarget, &'static PostProcessSettings, ); fn run( &self, _graph: &mut RenderGraphContext, render_context: &mut RenderContext, (view_target, _post_process_settings): QueryItem<Self::ViewQuery>, world: &World, ) -> Result<(), NodeRunError> { // ... } } // After impl ViewNode for PostProcessNode { type ViewData = ( &'static ViewTarget, &'static PostProcessSettings, ); fn run( &self, _graph: &mut RenderGraphContext, render_context: &mut RenderContext, (view_target, _post_process_settings): QueryItem<Self::ViewData>, world: &World, ) -> Result<(), NodeRunError> { // ... } } // Before impl<P: CachedRenderPipelinePhaseItem> RenderCommand<P> for SetItemPipeline { type Param = SRes<PipelineCache>; type ViewWorldQuery = (); type ItemWorldQuery = (); #[inline] fn render<'w>( item: &P, _view: (), _entity: (), pipeline_cache: SystemParamItem<'w, '_, Self::Param>, pass: &mut TrackedRenderPass<'w>, ) -> RenderCommandResult { // ... } } // After impl<P: CachedRenderPipelinePhaseItem> RenderCommand<P> for SetItemPipeline { type Param = SRes<PipelineCache>; type ViewData = (); type ItemData = (); #[inline] fn render<'w>( item: &P, _view: (), _entity: (), pipeline_cache: SystemParamItem<'w, '_, Self::Param>, pass: &mut TrackedRenderPass<'w>, ) -> RenderCommandResult { // ... } } ```
2023-12-12 19:45:50 +00:00
query::{Added, Changed, FilteredAccess, QueryFilter, With, Without},
Make `Resource` trait opt-in, requiring `#[derive(Resource)]` V2 (#5577) *This PR description is an edited copy of #5007, written by @alice-i-cecile.* # Objective Follow-up to https://github.com/bevyengine/bevy/pull/2254. The `Resource` trait currently has a blanket implementation for all types that meet its bounds. While ergonomic, this results in several drawbacks: * it is possible to make confusing, silent mistakes such as inserting a function pointer (Foo) rather than a value (Foo::Bar) as a resource * it is challenging to discover if a type is intended to be used as a resource * we cannot later add customization options (see the [RFC](https://github.com/bevyengine/rfcs/blob/main/rfcs/27-derive-component.md) for the equivalent choice for Component). * dependencies can use the same Rust type as a resource in invisibly conflicting ways * raw Rust types used as resources cannot preserve privacy appropriately, as anyone able to access that type can read and write to internal values * we cannot capture a definitive list of possible resources to display to users in an editor ## Notes to reviewers * Review this commit-by-commit; there's effectively no back-tracking and there's a lot of churn in some of these commits. *ira: My commits are not as well organized :')* * I've relaxed the bound on Local to Send + Sync + 'static: I don't think these concerns apply there, so this can keep things simple. Storing e.g. a u32 in a Local is fine, because there's a variable name attached explaining what it does. * I think this is a bad place for the Resource trait to live, but I've left it in place to make reviewing easier. IMO that's best tackled with https://github.com/bevyengine/bevy/issues/4981. ## Changelog `Resource` is no longer automatically implemented for all matching types. Instead, use the new `#[derive(Resource)]` macro. ## Migration Guide Add `#[derive(Resource)]` to all types you are using as a resource. If you are using a third party type as a resource, wrap it in a tuple struct to bypass orphan rules. Consider deriving `Deref` and `DerefMut` to improve ergonomics. `ClearColor` no longer implements `Component`. Using `ClearColor` as a component in 0.8 did nothing. Use the `ClearColorConfig` in the `Camera3d` and `Camera2d` components instead. Co-authored-by: Alice <alice.i.cecile@gmail.com> Co-authored-by: Alice Cecile <alice.i.cecile@gmail.com> Co-authored-by: devil-ira <justthecooldude@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-08-08 21:36:35 +00:00
system::Resource,
Implement WorldQuery for EntityRef (#6960) # Objective Partially address #5504. Fix #4278. Provide "whole entity" access in queries. This can be useful when you don't know at compile time what you're accessing (i.e. reflection via `ReflectComponent`). ## Solution Implement `WorldQuery` for `EntityRef`. - This provides read-only access to the entire entity, and supports anything that `EntityRef` can normally do. - It matches all archetypes and tables and will densely iterate when possible. - It marks all of the ArchetypeComponentIds of a matched archetype as read. - Adding it to a query will cause it to panic if used in conjunction with any other mutable access. - Expanded the docs on Query to advertise this feature. - Added tests to ensure the panics were working as intended. - Added `EntityRef` to the ECS prelude. To make this safe, `EntityRef::world` was removed as it gave potential `UnsafeCell`-like access to other parts of the `World` including aliased mutable access to the components it would otherwise read safely. ## Performance Not great beyond the additional parallelization opportunity over exclusive systems. The `EntityRef` is fetched from `Entities` like any other call to `World::entity`, which can be very random access heavy. This could be simplified if `ArchetypeRow` is available in `WorldQuery::fetch`'s arguments, but that's likely not something we should optimize for. ## Future work An equivalent API where it gives mutable access to all components on a entity can be done with a scoped version of `EntityMut` where it does not provide `&mut World` access nor allow for structural changes to the entity is feasible as well. This could be done as a safe alternative to exclusive system when structural mutation isn't required or the target set of entities is scoped. --- ## Changelog Added: `Access::has_any_write` Added: `EntityRef` now implements `WorldQuery`. Allows read-only access to the entire entity, incompatible with any other mutable access, can be mixed with `With`/`Without` filters for more targeted use. Added: `EntityRef` to `bevy::ecs::prelude`. Removed: `EntityRef::world` ## Migration Guide TODO --------- Co-authored-by: Carter Weinberg <weinbergcarter@gmail.com> Co-authored-by: Jakob Hellermann <jakob.hellermann@protonmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2023-06-22 21:20:00 +00:00
world::{EntityRef, Mut, World},
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
};
Remove task_pool parameter from par_for_each(_mut) (#4705) # Objective Fixes #3183. Requiring a `&TaskPool` parameter is sort of meaningless if the only correct one is to use the one provided by `Res<ComputeTaskPool>` all the time. ## Solution Have `QueryState` save a clone of the `ComputeTaskPool` which is used for all `par_for_each` functions. ~~Adds a small overhead of the internal `Arc` clone as a part of the startup, but the ergonomics win should be well worth this hardly-noticable overhead.~~ Updated the docs to note that it will panic the task pool is not present as a resource. # Future Work If https://github.com/bevyengine/rfcs/pull/54 is approved, we can replace these resource lookups with a static function call instead to get the `ComputeTaskPool`. --- ## Changelog Removed: The `task_pool` parameter of `Query(State)::par_for_each(_mut)`. These calls will use the `World`'s `ComputeTaskPool` resource instead. ## Migration Guide The `task_pool` parameter for `Query(State)::par_for_each(_mut)` has been removed. Remove these parameters from all calls to these functions. Before: ```rust fn parallel_system( task_pool: Res<ComputeTaskPool>, query: Query<&MyComponent>, ) { query.par_for_each(&task_pool, 32, |comp| { ... }); } ``` After: ```rust fn parallel_system(query: Query<&MyComponent>) { query.par_for_each(32, |comp| { ... }); } ``` If using `Query(State)` outside of a system run by the scheduler, you may need to manually configure and initialize a `ComputeTaskPool` as a resource in the `World`.
2022-05-30 16:59:38 +00:00
use bevy_tasks::{ComputeTaskPool, TaskPool};
use std::{
any::TypeId,
marker::PhantomData,
sync::{
atomic::{AtomicUsize, Ordering},
Arc, Mutex,
},
};
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
Make `Resource` trait opt-in, requiring `#[derive(Resource)]` V2 (#5577) *This PR description is an edited copy of #5007, written by @alice-i-cecile.* # Objective Follow-up to https://github.com/bevyengine/bevy/pull/2254. The `Resource` trait currently has a blanket implementation for all types that meet its bounds. While ergonomic, this results in several drawbacks: * it is possible to make confusing, silent mistakes such as inserting a function pointer (Foo) rather than a value (Foo::Bar) as a resource * it is challenging to discover if a type is intended to be used as a resource * we cannot later add customization options (see the [RFC](https://github.com/bevyengine/rfcs/blob/main/rfcs/27-derive-component.md) for the equivalent choice for Component). * dependencies can use the same Rust type as a resource in invisibly conflicting ways * raw Rust types used as resources cannot preserve privacy appropriately, as anyone able to access that type can read and write to internal values * we cannot capture a definitive list of possible resources to display to users in an editor ## Notes to reviewers * Review this commit-by-commit; there's effectively no back-tracking and there's a lot of churn in some of these commits. *ira: My commits are not as well organized :')* * I've relaxed the bound on Local to Send + Sync + 'static: I don't think these concerns apply there, so this can keep things simple. Storing e.g. a u32 in a Local is fine, because there's a variable name attached explaining what it does. * I think this is a bad place for the Resource trait to live, but I've left it in place to make reviewing easier. IMO that's best tackled with https://github.com/bevyengine/bevy/issues/4981. ## Changelog `Resource` is no longer automatically implemented for all matching types. Instead, use the new `#[derive(Resource)]` macro. ## Migration Guide Add `#[derive(Resource)]` to all types you are using as a resource. If you are using a third party type as a resource, wrap it in a tuple struct to bypass orphan rules. Consider deriving `Deref` and `DerefMut` to improve ergonomics. `ClearColor` no longer implements `Component`. Using `ClearColor` as a component in 0.8 did nothing. Use the `ClearColorConfig` in the `Camera3d` and `Camera2d` components instead. Co-authored-by: Alice <alice.i.cecile@gmail.com> Co-authored-by: Alice Cecile <alice.i.cecile@gmail.com> Co-authored-by: devil-ira <justthecooldude@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-08-08 21:36:35 +00:00
#[derive(Component, Resource, Debug, PartialEq, Eq, Clone, Copy)]
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
struct A(usize);
#[derive(Component, Debug, PartialEq, Eq, Clone, Copy)]
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
struct B(usize);
#[derive(Component, Debug, PartialEq, Eq, Clone, Copy)]
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
struct C;
#[derive(Default)]
struct NonSendA(usize, PhantomData<*mut ()>);
#[derive(Component, Clone, Debug)]
struct DropCk(Arc<AtomicUsize>);
impl DropCk {
fn new_pair() -> (Self, Arc<AtomicUsize>) {
let atomic = Arc::new(AtomicUsize::new(0));
(DropCk(atomic.clone()), atomic)
}
}
impl Drop for DropCk {
fn drop(&mut self) {
self.0.as_ref().fetch_add(1, Ordering::Relaxed);
}
}
#[derive(Component, Clone, Debug)]
#[component(storage = "SparseSet")]
struct DropCkSparse(DropCk);
#[derive(Component, Copy, Clone, PartialEq, Eq, Debug)]
#[component(storage = "Table")]
struct TableStored(&'static str);
#[derive(Component, Copy, Clone, PartialEq, Eq, Debug)]
#[component(storage = "SparseSet")]
struct SparseStored(u32);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
#[test]
fn random_access() {
let mut world = World::new();
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
let e = world.spawn((TableStored("abc"), SparseStored(123))).id();
let f = world
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
.spawn((TableStored("def"), SparseStored(456), A(1)))
.id();
assert_eq!(world.get::<TableStored>(e).unwrap().0, "abc");
assert_eq!(world.get::<SparseStored>(e).unwrap().0, 123);
assert_eq!(world.get::<TableStored>(f).unwrap().0, "def");
assert_eq!(world.get::<SparseStored>(f).unwrap().0, 456);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
// test archetype get_mut()
world.get_mut::<TableStored>(e).unwrap().0 = "xyz";
assert_eq!(world.get::<TableStored>(e).unwrap().0, "xyz");
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
// test sparse set get_mut()
world.get_mut::<SparseStored>(f).unwrap().0 = 42;
assert_eq!(world.get::<SparseStored>(f).unwrap().0, 42);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
}
#[test]
fn bundle_derive() {
let mut world = World::new();
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
#[derive(Bundle, PartialEq, Debug)]
Accept Bundles for insert and remove. Deprecate insert/remove_bundle (#6039) # Objective Take advantage of the "impl Bundle for Component" changes in #2975 / add the follow up changes discussed there. ## Solution - Change `insert` and `remove` to accept a Bundle instead of a Component (for both Commands and World) - Deprecate `insert_bundle`, `remove_bundle`, and `remove_bundle_intersection` - Add `remove_intersection` --- ## Changelog - Change `insert` and `remove` now accept a Bundle instead of a Component (for both Commands and World) - `insert_bundle` and `remove_bundle` are deprecated ## Migration Guide Replace `insert_bundle` with `insert`: ```rust // Old (0.8) commands.spawn().insert_bundle(SomeBundle::default()); // New (0.9) commands.spawn().insert(SomeBundle::default()); ``` Replace `remove_bundle` with `remove`: ```rust // Old (0.8) commands.entity(some_entity).remove_bundle::<SomeBundle>(); // New (0.9) commands.entity(some_entity).remove::<SomeBundle>(); ``` Replace `remove_bundle_intersection` with `remove_intersection`: ```rust // Old (0.8) world.entity_mut(some_entity).remove_bundle_intersection::<SomeBundle>(); // New (0.9) world.entity_mut(some_entity).remove_intersection::<SomeBundle>(); ``` Consider consolidating as many operations as possible to improve ergonomics and cut down on archetype moves: ```rust // Old (0.8) commands.spawn() .insert_bundle(SomeBundle::default()) .insert(SomeComponent); // New (0.9) - Option 1 commands.spawn().insert(( SomeBundle::default(), SomeComponent, )) // New (0.9) - Option 2 commands.spawn_bundle(( SomeBundle::default(), SomeComponent, )) ``` ## Next Steps Consider changing `spawn` to accept a bundle and deprecate `spawn_bundle`.
2022-09-21 21:47:53 +00:00
struct FooBundle {
x: TableStored,
y: SparseStored,
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
}
let mut ids = Vec::new();
Accept Bundles for insert and remove. Deprecate insert/remove_bundle (#6039) # Objective Take advantage of the "impl Bundle for Component" changes in #2975 / add the follow up changes discussed there. ## Solution - Change `insert` and `remove` to accept a Bundle instead of a Component (for both Commands and World) - Deprecate `insert_bundle`, `remove_bundle`, and `remove_bundle_intersection` - Add `remove_intersection` --- ## Changelog - Change `insert` and `remove` now accept a Bundle instead of a Component (for both Commands and World) - `insert_bundle` and `remove_bundle` are deprecated ## Migration Guide Replace `insert_bundle` with `insert`: ```rust // Old (0.8) commands.spawn().insert_bundle(SomeBundle::default()); // New (0.9) commands.spawn().insert(SomeBundle::default()); ``` Replace `remove_bundle` with `remove`: ```rust // Old (0.8) commands.entity(some_entity).remove_bundle::<SomeBundle>(); // New (0.9) commands.entity(some_entity).remove::<SomeBundle>(); ``` Replace `remove_bundle_intersection` with `remove_intersection`: ```rust // Old (0.8) world.entity_mut(some_entity).remove_bundle_intersection::<SomeBundle>(); // New (0.9) world.entity_mut(some_entity).remove_intersection::<SomeBundle>(); ``` Consider consolidating as many operations as possible to improve ergonomics and cut down on archetype moves: ```rust // Old (0.8) commands.spawn() .insert_bundle(SomeBundle::default()) .insert(SomeComponent); // New (0.9) - Option 1 commands.spawn().insert(( SomeBundle::default(), SomeComponent, )) // New (0.9) - Option 2 commands.spawn_bundle(( SomeBundle::default(), SomeComponent, )) ``` ## Next Steps Consider changing `spawn` to accept a bundle and deprecate `spawn_bundle`.
2022-09-21 21:47:53 +00:00
<FooBundle as Bundle>::component_ids(
&mut world.components,
&mut world.storages,
&mut |id| {
ids.push(id);
},
);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
assert_eq!(
ids,
&[
world.init_component::<TableStored>(),
world.init_component::<SparseStored>(),
]
);
let e1 = world
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
.spawn(FooBundle {
x: TableStored("abc"),
y: SparseStored(123),
})
.id();
let e2 = world
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
.spawn((TableStored("def"), SparseStored(456), A(1)))
.id();
assert_eq!(world.get::<TableStored>(e1).unwrap().0, "abc");
assert_eq!(world.get::<SparseStored>(e1).unwrap().0, 123);
assert_eq!(world.get::<TableStored>(e2).unwrap().0, "def");
assert_eq!(world.get::<SparseStored>(e2).unwrap().0, 456);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
// test archetype get_mut()
world.get_mut::<TableStored>(e1).unwrap().0 = "xyz";
assert_eq!(world.get::<TableStored>(e1).unwrap().0, "xyz");
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
// test sparse set get_mut()
world.get_mut::<SparseStored>(e2).unwrap().0 = 42;
assert_eq!(world.get::<SparseStored>(e2).unwrap().0, 42);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
assert_eq!(
world.entity_mut(e1).take::<FooBundle>().unwrap(),
Accept Bundles for insert and remove. Deprecate insert/remove_bundle (#6039) # Objective Take advantage of the "impl Bundle for Component" changes in #2975 / add the follow up changes discussed there. ## Solution - Change `insert` and `remove` to accept a Bundle instead of a Component (for both Commands and World) - Deprecate `insert_bundle`, `remove_bundle`, and `remove_bundle_intersection` - Add `remove_intersection` --- ## Changelog - Change `insert` and `remove` now accept a Bundle instead of a Component (for both Commands and World) - `insert_bundle` and `remove_bundle` are deprecated ## Migration Guide Replace `insert_bundle` with `insert`: ```rust // Old (0.8) commands.spawn().insert_bundle(SomeBundle::default()); // New (0.9) commands.spawn().insert(SomeBundle::default()); ``` Replace `remove_bundle` with `remove`: ```rust // Old (0.8) commands.entity(some_entity).remove_bundle::<SomeBundle>(); // New (0.9) commands.entity(some_entity).remove::<SomeBundle>(); ``` Replace `remove_bundle_intersection` with `remove_intersection`: ```rust // Old (0.8) world.entity_mut(some_entity).remove_bundle_intersection::<SomeBundle>(); // New (0.9) world.entity_mut(some_entity).remove_intersection::<SomeBundle>(); ``` Consider consolidating as many operations as possible to improve ergonomics and cut down on archetype moves: ```rust // Old (0.8) commands.spawn() .insert_bundle(SomeBundle::default()) .insert(SomeComponent); // New (0.9) - Option 1 commands.spawn().insert(( SomeBundle::default(), SomeComponent, )) // New (0.9) - Option 2 commands.spawn_bundle(( SomeBundle::default(), SomeComponent, )) ``` ## Next Steps Consider changing `spawn` to accept a bundle and deprecate `spawn_bundle`.
2022-09-21 21:47:53 +00:00
FooBundle {
x: TableStored("xyz"),
y: SparseStored(123),
}
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
);
#[derive(Bundle, PartialEq, Debug)]
Accept Bundles for insert and remove. Deprecate insert/remove_bundle (#6039) # Objective Take advantage of the "impl Bundle for Component" changes in #2975 / add the follow up changes discussed there. ## Solution - Change `insert` and `remove` to accept a Bundle instead of a Component (for both Commands and World) - Deprecate `insert_bundle`, `remove_bundle`, and `remove_bundle_intersection` - Add `remove_intersection` --- ## Changelog - Change `insert` and `remove` now accept a Bundle instead of a Component (for both Commands and World) - `insert_bundle` and `remove_bundle` are deprecated ## Migration Guide Replace `insert_bundle` with `insert`: ```rust // Old (0.8) commands.spawn().insert_bundle(SomeBundle::default()); // New (0.9) commands.spawn().insert(SomeBundle::default()); ``` Replace `remove_bundle` with `remove`: ```rust // Old (0.8) commands.entity(some_entity).remove_bundle::<SomeBundle>(); // New (0.9) commands.entity(some_entity).remove::<SomeBundle>(); ``` Replace `remove_bundle_intersection` with `remove_intersection`: ```rust // Old (0.8) world.entity_mut(some_entity).remove_bundle_intersection::<SomeBundle>(); // New (0.9) world.entity_mut(some_entity).remove_intersection::<SomeBundle>(); ``` Consider consolidating as many operations as possible to improve ergonomics and cut down on archetype moves: ```rust // Old (0.8) commands.spawn() .insert_bundle(SomeBundle::default()) .insert(SomeComponent); // New (0.9) - Option 1 commands.spawn().insert(( SomeBundle::default(), SomeComponent, )) // New (0.9) - Option 2 commands.spawn_bundle(( SomeBundle::default(), SomeComponent, )) ``` ## Next Steps Consider changing `spawn` to accept a bundle and deprecate `spawn_bundle`.
2022-09-21 21:47:53 +00:00
struct NestedBundle {
a: A,
Accept Bundles for insert and remove. Deprecate insert/remove_bundle (#6039) # Objective Take advantage of the "impl Bundle for Component" changes in #2975 / add the follow up changes discussed there. ## Solution - Change `insert` and `remove` to accept a Bundle instead of a Component (for both Commands and World) - Deprecate `insert_bundle`, `remove_bundle`, and `remove_bundle_intersection` - Add `remove_intersection` --- ## Changelog - Change `insert` and `remove` now accept a Bundle instead of a Component (for both Commands and World) - `insert_bundle` and `remove_bundle` are deprecated ## Migration Guide Replace `insert_bundle` with `insert`: ```rust // Old (0.8) commands.spawn().insert_bundle(SomeBundle::default()); // New (0.9) commands.spawn().insert(SomeBundle::default()); ``` Replace `remove_bundle` with `remove`: ```rust // Old (0.8) commands.entity(some_entity).remove_bundle::<SomeBundle>(); // New (0.9) commands.entity(some_entity).remove::<SomeBundle>(); ``` Replace `remove_bundle_intersection` with `remove_intersection`: ```rust // Old (0.8) world.entity_mut(some_entity).remove_bundle_intersection::<SomeBundle>(); // New (0.9) world.entity_mut(some_entity).remove_intersection::<SomeBundle>(); ``` Consider consolidating as many operations as possible to improve ergonomics and cut down on archetype moves: ```rust // Old (0.8) commands.spawn() .insert_bundle(SomeBundle::default()) .insert(SomeComponent); // New (0.9) - Option 1 commands.spawn().insert(( SomeBundle::default(), SomeComponent, )) // New (0.9) - Option 2 commands.spawn_bundle(( SomeBundle::default(), SomeComponent, )) ``` ## Next Steps Consider changing `spawn` to accept a bundle and deprecate `spawn_bundle`.
2022-09-21 21:47:53 +00:00
foo: FooBundle,
b: B,
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
}
let mut ids = Vec::new();
Accept Bundles for insert and remove. Deprecate insert/remove_bundle (#6039) # Objective Take advantage of the "impl Bundle for Component" changes in #2975 / add the follow up changes discussed there. ## Solution - Change `insert` and `remove` to accept a Bundle instead of a Component (for both Commands and World) - Deprecate `insert_bundle`, `remove_bundle`, and `remove_bundle_intersection` - Add `remove_intersection` --- ## Changelog - Change `insert` and `remove` now accept a Bundle instead of a Component (for both Commands and World) - `insert_bundle` and `remove_bundle` are deprecated ## Migration Guide Replace `insert_bundle` with `insert`: ```rust // Old (0.8) commands.spawn().insert_bundle(SomeBundle::default()); // New (0.9) commands.spawn().insert(SomeBundle::default()); ``` Replace `remove_bundle` with `remove`: ```rust // Old (0.8) commands.entity(some_entity).remove_bundle::<SomeBundle>(); // New (0.9) commands.entity(some_entity).remove::<SomeBundle>(); ``` Replace `remove_bundle_intersection` with `remove_intersection`: ```rust // Old (0.8) world.entity_mut(some_entity).remove_bundle_intersection::<SomeBundle>(); // New (0.9) world.entity_mut(some_entity).remove_intersection::<SomeBundle>(); ``` Consider consolidating as many operations as possible to improve ergonomics and cut down on archetype moves: ```rust // Old (0.8) commands.spawn() .insert_bundle(SomeBundle::default()) .insert(SomeComponent); // New (0.9) - Option 1 commands.spawn().insert(( SomeBundle::default(), SomeComponent, )) // New (0.9) - Option 2 commands.spawn_bundle(( SomeBundle::default(), SomeComponent, )) ``` ## Next Steps Consider changing `spawn` to accept a bundle and deprecate `spawn_bundle`.
2022-09-21 21:47:53 +00:00
<NestedBundle as Bundle>::component_ids(
&mut world.components,
&mut world.storages,
&mut |id| {
ids.push(id);
},
);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
assert_eq!(
ids,
&[
world.init_component::<A>(),
world.init_component::<TableStored>(),
world.init_component::<SparseStored>(),
world.init_component::<B>(),
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
]
);
let e3 = world
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
.spawn(NestedBundle {
a: A(1),
Accept Bundles for insert and remove. Deprecate insert/remove_bundle (#6039) # Objective Take advantage of the "impl Bundle for Component" changes in #2975 / add the follow up changes discussed there. ## Solution - Change `insert` and `remove` to accept a Bundle instead of a Component (for both Commands and World) - Deprecate `insert_bundle`, `remove_bundle`, and `remove_bundle_intersection` - Add `remove_intersection` --- ## Changelog - Change `insert` and `remove` now accept a Bundle instead of a Component (for both Commands and World) - `insert_bundle` and `remove_bundle` are deprecated ## Migration Guide Replace `insert_bundle` with `insert`: ```rust // Old (0.8) commands.spawn().insert_bundle(SomeBundle::default()); // New (0.9) commands.spawn().insert(SomeBundle::default()); ``` Replace `remove_bundle` with `remove`: ```rust // Old (0.8) commands.entity(some_entity).remove_bundle::<SomeBundle>(); // New (0.9) commands.entity(some_entity).remove::<SomeBundle>(); ``` Replace `remove_bundle_intersection` with `remove_intersection`: ```rust // Old (0.8) world.entity_mut(some_entity).remove_bundle_intersection::<SomeBundle>(); // New (0.9) world.entity_mut(some_entity).remove_intersection::<SomeBundle>(); ``` Consider consolidating as many operations as possible to improve ergonomics and cut down on archetype moves: ```rust // Old (0.8) commands.spawn() .insert_bundle(SomeBundle::default()) .insert(SomeComponent); // New (0.9) - Option 1 commands.spawn().insert(( SomeBundle::default(), SomeComponent, )) // New (0.9) - Option 2 commands.spawn_bundle(( SomeBundle::default(), SomeComponent, )) ``` ## Next Steps Consider changing `spawn` to accept a bundle and deprecate `spawn_bundle`.
2022-09-21 21:47:53 +00:00
foo: FooBundle {
x: TableStored("ghi"),
y: SparseStored(789),
},
b: B(2),
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
})
.id();
assert_eq!(world.get::<TableStored>(e3).unwrap().0, "ghi");
assert_eq!(world.get::<SparseStored>(e3).unwrap().0, 789);
assert_eq!(world.get::<A>(e3).unwrap().0, 1);
assert_eq!(world.get::<B>(e3).unwrap().0, 2);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
assert_eq!(
world.entity_mut(e3).take::<NestedBundle>().unwrap(),
Accept Bundles for insert and remove. Deprecate insert/remove_bundle (#6039) # Objective Take advantage of the "impl Bundle for Component" changes in #2975 / add the follow up changes discussed there. ## Solution - Change `insert` and `remove` to accept a Bundle instead of a Component (for both Commands and World) - Deprecate `insert_bundle`, `remove_bundle`, and `remove_bundle_intersection` - Add `remove_intersection` --- ## Changelog - Change `insert` and `remove` now accept a Bundle instead of a Component (for both Commands and World) - `insert_bundle` and `remove_bundle` are deprecated ## Migration Guide Replace `insert_bundle` with `insert`: ```rust // Old (0.8) commands.spawn().insert_bundle(SomeBundle::default()); // New (0.9) commands.spawn().insert(SomeBundle::default()); ``` Replace `remove_bundle` with `remove`: ```rust // Old (0.8) commands.entity(some_entity).remove_bundle::<SomeBundle>(); // New (0.9) commands.entity(some_entity).remove::<SomeBundle>(); ``` Replace `remove_bundle_intersection` with `remove_intersection`: ```rust // Old (0.8) world.entity_mut(some_entity).remove_bundle_intersection::<SomeBundle>(); // New (0.9) world.entity_mut(some_entity).remove_intersection::<SomeBundle>(); ``` Consider consolidating as many operations as possible to improve ergonomics and cut down on archetype moves: ```rust // Old (0.8) commands.spawn() .insert_bundle(SomeBundle::default()) .insert(SomeComponent); // New (0.9) - Option 1 commands.spawn().insert(( SomeBundle::default(), SomeComponent, )) // New (0.9) - Option 2 commands.spawn_bundle(( SomeBundle::default(), SomeComponent, )) ``` ## Next Steps Consider changing `spawn` to accept a bundle and deprecate `spawn_bundle`.
2022-09-21 21:47:53 +00:00
NestedBundle {
a: A(1),
Accept Bundles for insert and remove. Deprecate insert/remove_bundle (#6039) # Objective Take advantage of the "impl Bundle for Component" changes in #2975 / add the follow up changes discussed there. ## Solution - Change `insert` and `remove` to accept a Bundle instead of a Component (for both Commands and World) - Deprecate `insert_bundle`, `remove_bundle`, and `remove_bundle_intersection` - Add `remove_intersection` --- ## Changelog - Change `insert` and `remove` now accept a Bundle instead of a Component (for both Commands and World) - `insert_bundle` and `remove_bundle` are deprecated ## Migration Guide Replace `insert_bundle` with `insert`: ```rust // Old (0.8) commands.spawn().insert_bundle(SomeBundle::default()); // New (0.9) commands.spawn().insert(SomeBundle::default()); ``` Replace `remove_bundle` with `remove`: ```rust // Old (0.8) commands.entity(some_entity).remove_bundle::<SomeBundle>(); // New (0.9) commands.entity(some_entity).remove::<SomeBundle>(); ``` Replace `remove_bundle_intersection` with `remove_intersection`: ```rust // Old (0.8) world.entity_mut(some_entity).remove_bundle_intersection::<SomeBundle>(); // New (0.9) world.entity_mut(some_entity).remove_intersection::<SomeBundle>(); ``` Consider consolidating as many operations as possible to improve ergonomics and cut down on archetype moves: ```rust // Old (0.8) commands.spawn() .insert_bundle(SomeBundle::default()) .insert(SomeComponent); // New (0.9) - Option 1 commands.spawn().insert(( SomeBundle::default(), SomeComponent, )) // New (0.9) - Option 2 commands.spawn_bundle(( SomeBundle::default(), SomeComponent, )) ``` ## Next Steps Consider changing `spawn` to accept a bundle and deprecate `spawn_bundle`.
2022-09-21 21:47:53 +00:00
foo: FooBundle {
x: TableStored("ghi"),
y: SparseStored(789),
},
b: B(2),
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
}
);
#[derive(Default, Component, PartialEq, Debug)]
struct Ignored;
#[derive(Bundle, PartialEq, Debug)]
struct BundleWithIgnored {
c: C,
#[bundle(ignore)]
ignored: Ignored,
}
let mut ids = Vec::new();
<BundleWithIgnored as Bundle>::component_ids(
&mut world.components,
&mut world.storages,
&mut |id| {
ids.push(id);
},
);
assert_eq!(ids, &[world.init_component::<C>(),]);
let e4 = world
.spawn(BundleWithIgnored {
c: C,
ignored: Ignored,
})
.id();
assert_eq!(world.get::<C>(e4).unwrap(), &C);
assert_eq!(world.get::<Ignored>(e4), None);
assert_eq!(
world.entity_mut(e4).take::<BundleWithIgnored>().unwrap(),
BundleWithIgnored {
c: C,
ignored: Ignored,
}
);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
}
#[test]
fn despawn_table_storage() {
let mut world = World::new();
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
let e = world.spawn((TableStored("abc"), A(123))).id();
let f = world.spawn((TableStored("def"), A(456))).id();
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
assert_eq!(world.entities.len(), 2);
assert!(world.despawn(e));
assert_eq!(world.entities.len(), 1);
assert!(world.get::<TableStored>(e).is_none());
assert!(world.get::<A>(e).is_none());
assert_eq!(world.get::<TableStored>(f).unwrap().0, "def");
assert_eq!(world.get::<A>(f).unwrap().0, 456);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
}
#[test]
fn despawn_mixed_storage() {
let mut world = World::new();
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
let e = world.spawn((TableStored("abc"), SparseStored(123))).id();
let f = world.spawn((TableStored("def"), SparseStored(456))).id();
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
assert_eq!(world.entities.len(), 2);
assert!(world.despawn(e));
assert_eq!(world.entities.len(), 1);
assert!(world.get::<TableStored>(e).is_none());
assert!(world.get::<SparseStored>(e).is_none());
assert_eq!(world.get::<TableStored>(f).unwrap().0, "def");
assert_eq!(world.get::<SparseStored>(f).unwrap().0, 456);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
}
#[test]
fn query_all() {
let mut world = World::new();
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
let e = world.spawn((TableStored("abc"), A(123))).id();
let f = world.spawn((TableStored("def"), A(456))).id();
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
let ents = world
.query::<(Entity, &A, &TableStored)>()
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
.iter(&world)
.map(|(e, &i, &s)| (e, i, s))
.collect::<Vec<_>>();
assert_eq!(
ents,
&[
(e, A(123), TableStored("abc")),
(f, A(456), TableStored("def"))
]
);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
}
#[test]
fn query_all_for_each() {
let mut world = World::new();
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
let e = world.spawn((TableStored("abc"), A(123))).id();
let f = world.spawn((TableStored("def"), A(456))).id();
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
let mut results = Vec::new();
world
.query::<(Entity, &A, &TableStored)>()
Override QueryIter::fold to port Query::for_each perf gains to select Iterator combinators (#6773) # Objective After #6547, `Query::for_each` has been capable of automatic vectorization on certain queries, which is seeing a notable (>50% CPU time improvements) for iteration. However, `Query::for_each` isn't idiomatic Rust, and lacks the flexibility of iterator combinators. Ideally, `Query::iter` and friends should be able to achieve the same results. However, this does seem to blocked upstream (rust-lang/rust#104914) by Rust's loop optimizations. ## Solution This is an intermediate solution and refactor. This moves the `Query::for_each` implementation onto the `Iterator::fold` implementation for `QueryIter` instead. This should result in the same automatic vectorization optimization on all `Iterator` functions that internally use fold, including `Iterator::for_each`, `Iterator::count`, etc. With this, it should close the gap between the two completely. Internally, this PR changes `Query::for_each` to use `query.iter().for_each(..)` instead of the duplicated implementation. Separately, the duplicate implementations of internal iteration (i.e. `Query::par_for_each`) now use portions of the current `Query::for_each` implementation factored out into their own functions. This also massively cleans up our internal fragmentation of internal iteration options, deduplicating the iteration code used in `for_each` and `par_iter().for_each()`. --- ## Changelog Changed: `Query::for_each`, `Query::for_each_mut`, `Query::for_each`, and `Query::for_each_mut` have been moved to `QueryIter`'s `Iterator::for_each` implementation, and still retains their performance improvements over normal iteration. These APIs are deprecated in 0.13 and will be removed in 0.14. --------- Co-authored-by: JoJoJet <21144246+JoJoJet@users.noreply.github.com> Co-authored-by: Alice Cecile <alice.i.cecile@gmail.com>
2023-12-01 09:09:55 +00:00
.iter(&world)
.for_each(|(e, &i, &s)| results.push((e, i, s)));
assert_eq!(
results,
&[
(e, A(123), TableStored("abc")),
(f, A(456), TableStored("def"))
]
);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
}
#[test]
fn query_single_component() {
let mut world = World::new();
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
let e = world.spawn((TableStored("abc"), A(123))).id();
let f = world.spawn((TableStored("def"), A(456), B(1))).id();
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
let ents = world
.query::<(Entity, &A)>()
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
.iter(&world)
.map(|(e, &i)| (e, i))
.collect::<Vec<_>>();
assert_eq!(ents, &[(e, A(123)), (f, A(456))]);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
}
#[test]
fn stateful_query_handles_new_archetype() {
let mut world = World::new();
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
let e = world.spawn((TableStored("abc"), A(123))).id();
let mut query = world.query::<(Entity, &A)>();
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
let ents = query.iter(&world).map(|(e, &i)| (e, i)).collect::<Vec<_>>();
assert_eq!(ents, &[(e, A(123))]);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
let f = world.spawn((TableStored("def"), A(456), B(1))).id();
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
let ents = query.iter(&world).map(|(e, &i)| (e, i)).collect::<Vec<_>>();
assert_eq!(ents, &[(e, A(123)), (f, A(456))]);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
}
#[test]
fn query_single_component_for_each() {
let mut world = World::new();
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
let e = world.spawn((TableStored("abc"), A(123))).id();
let f = world.spawn((TableStored("def"), A(456), B(1))).id();
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
let mut results = Vec::new();
world
.query::<(Entity, &A)>()
Override QueryIter::fold to port Query::for_each perf gains to select Iterator combinators (#6773) # Objective After #6547, `Query::for_each` has been capable of automatic vectorization on certain queries, which is seeing a notable (>50% CPU time improvements) for iteration. However, `Query::for_each` isn't idiomatic Rust, and lacks the flexibility of iterator combinators. Ideally, `Query::iter` and friends should be able to achieve the same results. However, this does seem to blocked upstream (rust-lang/rust#104914) by Rust's loop optimizations. ## Solution This is an intermediate solution and refactor. This moves the `Query::for_each` implementation onto the `Iterator::fold` implementation for `QueryIter` instead. This should result in the same automatic vectorization optimization on all `Iterator` functions that internally use fold, including `Iterator::for_each`, `Iterator::count`, etc. With this, it should close the gap between the two completely. Internally, this PR changes `Query::for_each` to use `query.iter().for_each(..)` instead of the duplicated implementation. Separately, the duplicate implementations of internal iteration (i.e. `Query::par_for_each`) now use portions of the current `Query::for_each` implementation factored out into their own functions. This also massively cleans up our internal fragmentation of internal iteration options, deduplicating the iteration code used in `for_each` and `par_iter().for_each()`. --- ## Changelog Changed: `Query::for_each`, `Query::for_each_mut`, `Query::for_each`, and `Query::for_each_mut` have been moved to `QueryIter`'s `Iterator::for_each` implementation, and still retains their performance improvements over normal iteration. These APIs are deprecated in 0.13 and will be removed in 0.14. --------- Co-authored-by: JoJoJet <21144246+JoJoJet@users.noreply.github.com> Co-authored-by: Alice Cecile <alice.i.cecile@gmail.com>
2023-12-01 09:09:55 +00:00
.iter(&world)
.for_each(|(e, &i)| results.push((e, i)));
assert_eq!(results, &[(e, A(123)), (f, A(456))]);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
}
#[test]
fn par_for_each_dense() {
ComputeTaskPool::get_or_init(TaskPool::default);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
let mut world = World::new();
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
let e1 = world.spawn(A(1)).id();
let e2 = world.spawn(A(2)).id();
let e3 = world.spawn(A(3)).id();
let e4 = world.spawn((A(4), B(1))).id();
let e5 = world.spawn((A(5), B(1))).id();
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
let results = Arc::new(Mutex::new(Vec::new()));
world
.query::<(Entity, &A)>()
Basic adaptive batching for parallel query iteration (#4777) # Objective Fixes #3184. Fixes #6640. Fixes #4798. Using `Query::par_for_each(_mut)` currently requires a `batch_size` parameter, which affects how it chunks up large archetypes and tables into smaller chunks to run in parallel. Tuning this value is difficult, as the performance characteristics entirely depends on the state of the `World` it's being run on. Typically, users will just use a flat constant and just tune it by hand until it performs well in some benchmarks. However, this is both error prone and risks overfitting the tuning on that benchmark. This PR proposes a naive automatic batch-size computation based on the current state of the `World`. ## Background `Query::par_for_each(_mut)` schedules a new Task for every archetype or table that it matches. Archetypes/tables larger than the batch size are chunked into smaller tasks. Assuming every entity matched by the query has an identical workload, this makes the worst case scenario involve using a batch size equal to the size of the largest matched archetype or table. Conversely, a batch size of `max {archetype, table} size / thread count * COUNT_PER_THREAD` is likely the sweetspot where the overhead of scheduling tasks is minimized, at least not without grouping small archetypes/tables together. There is also likely a strict minimum batch size below which the overhead of scheduling these tasks is heavier than running the entire thing single-threaded. ## Solution - [x] Remove the `batch_size` from `Query(State)::par_for_each` and friends. - [x] Add a check to compute `batch_size = max {archeytpe/table} size / thread count * COUNT_PER_THREAD` - [x] ~~Panic if thread count is 0.~~ Defer to `for_each` if the thread count is 1 or less. - [x] Early return if there is no matched table/archetype. - [x] Add override option for users have queries that strongly violate the initial assumption that all iterated entities have an equal workload. --- ## Changelog Changed: `Query::par_for_each(_mut)` has been changed to `Query::par_iter(_mut)` and will now automatically try to produce a batch size for callers based on the current `World` state. ## Migration Guide The `batch_size` parameter for `Query(State)::par_for_each(_mut)` has been removed. These calls will automatically compute a batch size for you. Remove these parameters from all calls to these functions. Before: ```rust fn parallel_system(query: Query<&MyComponent>) { query.par_for_each(32, |comp| { ... }); } ``` After: ```rust fn parallel_system(query: Query<&MyComponent>) { query.par_iter().for_each(|comp| { ... }); } ``` Co-authored-by: Arnav Choubey <56453634+x-52@users.noreply.github.com> Co-authored-by: Robert Swain <robert.swain@gmail.com> Co-authored-by: François <mockersf@gmail.com> Co-authored-by: Corey Farwell <coreyf@rwell.org> Co-authored-by: Aevyrie <aevyrie@gmail.com>
2023-01-20 08:47:20 +00:00
.par_iter(&world)
.for_each(|(e, &A(i))| {
results.lock().unwrap().push((e, i));
});
results.lock().unwrap().sort();
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
assert_eq!(
&*results.lock().unwrap(),
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
&[(e1, 1), (e2, 2), (e3, 3), (e4, 4), (e5, 5)]
);
}
#[test]
fn par_for_each_sparse() {
ComputeTaskPool::get_or_init(TaskPool::default);
let mut world = World::new();
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
let e1 = world.spawn(SparseStored(1)).id();
let e2 = world.spawn(SparseStored(2)).id();
let e3 = world.spawn(SparseStored(3)).id();
let e4 = world.spawn((SparseStored(4), A(1))).id();
let e5 = world.spawn((SparseStored(5), A(1))).id();
let results = Arc::new(Mutex::new(Vec::new()));
Basic adaptive batching for parallel query iteration (#4777) # Objective Fixes #3184. Fixes #6640. Fixes #4798. Using `Query::par_for_each(_mut)` currently requires a `batch_size` parameter, which affects how it chunks up large archetypes and tables into smaller chunks to run in parallel. Tuning this value is difficult, as the performance characteristics entirely depends on the state of the `World` it's being run on. Typically, users will just use a flat constant and just tune it by hand until it performs well in some benchmarks. However, this is both error prone and risks overfitting the tuning on that benchmark. This PR proposes a naive automatic batch-size computation based on the current state of the `World`. ## Background `Query::par_for_each(_mut)` schedules a new Task for every archetype or table that it matches. Archetypes/tables larger than the batch size are chunked into smaller tasks. Assuming every entity matched by the query has an identical workload, this makes the worst case scenario involve using a batch size equal to the size of the largest matched archetype or table. Conversely, a batch size of `max {archetype, table} size / thread count * COUNT_PER_THREAD` is likely the sweetspot where the overhead of scheduling tasks is minimized, at least not without grouping small archetypes/tables together. There is also likely a strict minimum batch size below which the overhead of scheduling these tasks is heavier than running the entire thing single-threaded. ## Solution - [x] Remove the `batch_size` from `Query(State)::par_for_each` and friends. - [x] Add a check to compute `batch_size = max {archeytpe/table} size / thread count * COUNT_PER_THREAD` - [x] ~~Panic if thread count is 0.~~ Defer to `for_each` if the thread count is 1 or less. - [x] Early return if there is no matched table/archetype. - [x] Add override option for users have queries that strongly violate the initial assumption that all iterated entities have an equal workload. --- ## Changelog Changed: `Query::par_for_each(_mut)` has been changed to `Query::par_iter(_mut)` and will now automatically try to produce a batch size for callers based on the current `World` state. ## Migration Guide The `batch_size` parameter for `Query(State)::par_for_each(_mut)` has been removed. These calls will automatically compute a batch size for you. Remove these parameters from all calls to these functions. Before: ```rust fn parallel_system(query: Query<&MyComponent>) { query.par_for_each(32, |comp| { ... }); } ``` After: ```rust fn parallel_system(query: Query<&MyComponent>) { query.par_iter().for_each(|comp| { ... }); } ``` Co-authored-by: Arnav Choubey <56453634+x-52@users.noreply.github.com> Co-authored-by: Robert Swain <robert.swain@gmail.com> Co-authored-by: François <mockersf@gmail.com> Co-authored-by: Corey Farwell <coreyf@rwell.org> Co-authored-by: Aevyrie <aevyrie@gmail.com>
2023-01-20 08:47:20 +00:00
world
.query::<(Entity, &SparseStored)>()
.par_iter(&world)
.for_each(|(e, &SparseStored(i))| results.lock().unwrap().push((e, i)));
results.lock().unwrap().sort();
assert_eq!(
&*results.lock().unwrap(),
&[(e1, 1), (e2, 2), (e3, 3), (e4, 4), (e5, 5)]
);
}
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
#[test]
fn query_missing_component() {
let mut world = World::new();
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
world.spawn((TableStored("abc"), A(123)));
world.spawn((TableStored("def"), A(456)));
assert!(world.query::<(&B, &A)>().iter(&world).next().is_none());
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
}
#[test]
fn query_sparse_component() {
let mut world = World::new();
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
world.spawn((TableStored("abc"), A(123)));
let f = world.spawn((TableStored("def"), A(456), B(1))).id();
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
let ents = world
.query::<(Entity, &B)>()
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
.iter(&world)
.map(|(e, &b)| (e, b))
.collect::<Vec<_>>();
assert_eq!(ents, &[(f, B(1))]);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
}
#[test]
fn query_filter_with() {
let mut world = World::new();
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
world.spawn((A(123), B(1)));
world.spawn(A(456));
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
let result = world
.query_filtered::<&A, With<B>>()
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
.iter(&world)
.cloned()
.collect::<Vec<_>>();
assert_eq!(result, vec![A(123)]);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
}
#[test]
fn query_filter_with_for_each() {
let mut world = World::new();
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
world.spawn((A(123), B(1)));
world.spawn(A(456));
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
let mut results = Vec::new();
world
.query_filtered::<&A, With<B>>()
Override QueryIter::fold to port Query::for_each perf gains to select Iterator combinators (#6773) # Objective After #6547, `Query::for_each` has been capable of automatic vectorization on certain queries, which is seeing a notable (>50% CPU time improvements) for iteration. However, `Query::for_each` isn't idiomatic Rust, and lacks the flexibility of iterator combinators. Ideally, `Query::iter` and friends should be able to achieve the same results. However, this does seem to blocked upstream (rust-lang/rust#104914) by Rust's loop optimizations. ## Solution This is an intermediate solution and refactor. This moves the `Query::for_each` implementation onto the `Iterator::fold` implementation for `QueryIter` instead. This should result in the same automatic vectorization optimization on all `Iterator` functions that internally use fold, including `Iterator::for_each`, `Iterator::count`, etc. With this, it should close the gap between the two completely. Internally, this PR changes `Query::for_each` to use `query.iter().for_each(..)` instead of the duplicated implementation. Separately, the duplicate implementations of internal iteration (i.e. `Query::par_for_each`) now use portions of the current `Query::for_each` implementation factored out into their own functions. This also massively cleans up our internal fragmentation of internal iteration options, deduplicating the iteration code used in `for_each` and `par_iter().for_each()`. --- ## Changelog Changed: `Query::for_each`, `Query::for_each_mut`, `Query::for_each`, and `Query::for_each_mut` have been moved to `QueryIter`'s `Iterator::for_each` implementation, and still retains their performance improvements over normal iteration. These APIs are deprecated in 0.13 and will be removed in 0.14. --------- Co-authored-by: JoJoJet <21144246+JoJoJet@users.noreply.github.com> Co-authored-by: Alice Cecile <alice.i.cecile@gmail.com>
2023-12-01 09:09:55 +00:00
.iter(&world)
.for_each(|i| results.push(*i));
assert_eq!(results, vec![A(123)]);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
}
#[test]
fn query_filter_with_sparse() {
let mut world = World::new();
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
world.spawn((A(123), SparseStored(321)));
world.spawn(A(456));
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
let result = world
.query_filtered::<&A, With<SparseStored>>()
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
.iter(&world)
.cloned()
.collect::<Vec<_>>();
assert_eq!(result, vec![A(123)]);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
}
#[test]
fn query_filter_with_sparse_for_each() {
let mut world = World::new();
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
world.spawn((A(123), SparseStored(321)));
world.spawn(A(456));
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
let mut results = Vec::new();
world
.query_filtered::<&A, With<SparseStored>>()
Override QueryIter::fold to port Query::for_each perf gains to select Iterator combinators (#6773) # Objective After #6547, `Query::for_each` has been capable of automatic vectorization on certain queries, which is seeing a notable (>50% CPU time improvements) for iteration. However, `Query::for_each` isn't idiomatic Rust, and lacks the flexibility of iterator combinators. Ideally, `Query::iter` and friends should be able to achieve the same results. However, this does seem to blocked upstream (rust-lang/rust#104914) by Rust's loop optimizations. ## Solution This is an intermediate solution and refactor. This moves the `Query::for_each` implementation onto the `Iterator::fold` implementation for `QueryIter` instead. This should result in the same automatic vectorization optimization on all `Iterator` functions that internally use fold, including `Iterator::for_each`, `Iterator::count`, etc. With this, it should close the gap between the two completely. Internally, this PR changes `Query::for_each` to use `query.iter().for_each(..)` instead of the duplicated implementation. Separately, the duplicate implementations of internal iteration (i.e. `Query::par_for_each`) now use portions of the current `Query::for_each` implementation factored out into their own functions. This also massively cleans up our internal fragmentation of internal iteration options, deduplicating the iteration code used in `for_each` and `par_iter().for_each()`. --- ## Changelog Changed: `Query::for_each`, `Query::for_each_mut`, `Query::for_each`, and `Query::for_each_mut` have been moved to `QueryIter`'s `Iterator::for_each` implementation, and still retains their performance improvements over normal iteration. These APIs are deprecated in 0.13 and will be removed in 0.14. --------- Co-authored-by: JoJoJet <21144246+JoJoJet@users.noreply.github.com> Co-authored-by: Alice Cecile <alice.i.cecile@gmail.com>
2023-12-01 09:09:55 +00:00
.iter(&world)
.for_each(|i| results.push(*i));
assert_eq!(results, vec![A(123)]);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
}
#[test]
fn query_filter_without() {
let mut world = World::new();
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
world.spawn((A(123), B(321)));
world.spawn(A(456));
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
let result = world
.query_filtered::<&A, Without<B>>()
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
.iter(&world)
.cloned()
.collect::<Vec<_>>();
assert_eq!(result, vec![A(456)]);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
}
#[test]
fn query_optional_component_table() {
let mut world = World::new();
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
let e = world.spawn((TableStored("abc"), A(123))).id();
let f = world.spawn((TableStored("def"), A(456), B(1))).id();
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
// this should be skipped
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
world.spawn(TableStored("abc"));
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
let ents = world
.query::<(Entity, Option<&B>, &A)>()
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
.iter(&world)
.map(|(e, b, &i)| (e, b.copied(), i))
.collect::<Vec<_>>();
assert_eq!(ents, &[(e, None, A(123)), (f, Some(B(1)), A(456))]);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
}
#[test]
fn query_optional_component_sparse() {
let mut world = World::new();
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
let e = world.spawn((TableStored("abc"), A(123))).id();
let f = world
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
.spawn((TableStored("def"), A(456), SparseStored(1)))
.id();
// this should be skipped
// world.spawn(SparseStored(1));
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
let ents = world
.query::<(Entity, Option<&SparseStored>, &A)>()
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
.iter(&world)
.map(|(e, b, &i)| (e, b.copied(), i))
.collect::<Vec<_>>();
assert_eq!(
ents,
&[(e, None, A(123)), (f, Some(SparseStored(1)), A(456))]
);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
}
#[test]
fn query_optional_component_sparse_no_match() {
let mut world = World::new();
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
let e = world.spawn((TableStored("abc"), A(123))).id();
let f = world.spawn((TableStored("def"), A(456))).id();
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
// // this should be skipped
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
world.spawn(TableStored("abc"));
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
let ents = world
.query::<(Entity, Option<&SparseStored>, &A)>()
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
.iter(&world)
.map(|(e, b, &i)| (e, b.copied(), i))
.collect::<Vec<_>>();
assert_eq!(ents, &[(e, None, A(123)), (f, None, A(456))]);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
}
#[test]
fn add_remove_components() {
let mut world = World::new();
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
let e1 = world.spawn((A(1), B(3), TableStored("abc"))).id();
let e2 = world.spawn((A(2), B(4), TableStored("xyz"))).id();
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
assert_eq!(
world
.query::<(Entity, &A, &B)>()
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
.iter(&world)
.map(|(e, &i, &b)| (e, i, b))
.collect::<Vec<_>>(),
&[(e1, A(1), B(3)), (e2, A(2), B(4))]
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
);
assert_eq!(world.entity_mut(e1).take::<A>(), Some(A(1)));
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
assert_eq!(
world
.query::<(Entity, &A, &B)>()
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
.iter(&world)
.map(|(e, &i, &b)| (e, i, b))
.collect::<Vec<_>>(),
&[(e2, A(2), B(4))]
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
);
assert_eq!(
world
.query::<(Entity, &B, &TableStored)>()
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
.iter(&world)
.map(|(e, &B(b), &TableStored(s))| (e, b, s))
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
.collect::<Vec<_>>(),
&[(e2, 4, "xyz"), (e1, 3, "abc")]
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
);
world.entity_mut(e1).insert(A(43));
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
assert_eq!(
world
.query::<(Entity, &A, &B)>()
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
.iter(&world)
.map(|(e, &i, &b)| (e, i, b))
.collect::<Vec<_>>(),
&[(e2, A(2), B(4)), (e1, A(43), B(3))]
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
);
world.entity_mut(e1).insert(C);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
assert_eq!(
world
.query::<(Entity, &C)>()
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
.iter(&world)
.map(|(e, &f)| (e, f))
.collect::<Vec<_>>(),
&[(e1, C)]
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
);
}
#[test]
fn table_add_remove_many() {
let mut world = World::default();
#[cfg(miri)]
let (mut entities, to) = {
let to = 10;
(Vec::with_capacity(to), to)
};
#[cfg(not(miri))]
let (mut entities, to) = {
let to = 10_000;
(Vec::with_capacity(to), to)
};
for _ in 0..to {
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
entities.push(world.spawn(B(0)).id());
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
}
for (i, entity) in entities.iter().cloned().enumerate() {
world.entity_mut(entity).insert(A(i));
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
}
for (i, entity) in entities.iter().cloned().enumerate() {
assert_eq!(world.entity_mut(entity).take::<A>(), Some(A(i)));
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
}
}
#[test]
fn sparse_set_add_remove_many() {
let mut world = World::default();
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
let mut entities = Vec::with_capacity(1000);
for _ in 0..4 {
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
entities.push(world.spawn(A(2)).id());
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
}
for (i, entity) in entities.iter().cloned().enumerate() {
world.entity_mut(entity).insert(SparseStored(i as u32));
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
}
for (i, entity) in entities.iter().cloned().enumerate() {
assert_eq!(
world.entity_mut(entity).take::<SparseStored>(),
Some(SparseStored(i as u32))
);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
}
}
#[test]
fn remove_missing() {
let mut world = World::new();
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
let e = world.spawn((TableStored("abc"), A(123))).id();
assert!(world.entity_mut(e).take::<B>().is_none());
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
}
#[test]
fn spawn_batch() {
let mut world = World::new();
world.spawn_batch((0..100).map(|x| (A(x), TableStored("abc"))));
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
let values = world
.query::<&A>()
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
.iter(&world)
.map(|v| v.0)
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
.collect::<Vec<_>>();
let expected = (0..100).collect::<Vec<_>>();
assert_eq!(values, expected);
}
#[test]
fn query_get() {
let mut world = World::new();
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
let a = world.spawn((TableStored("abc"), A(123))).id();
let b = world.spawn((TableStored("def"), A(456))).id();
let c = world.spawn((TableStored("ghi"), A(789), B(1))).id();
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
let mut i32_query = world.query::<&A>();
assert_eq!(i32_query.get(&world, a).unwrap().0, 123);
assert_eq!(i32_query.get(&world, b).unwrap().0, 456);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
let mut i32_bool_query = world.query::<(&A, &B)>();
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
assert!(i32_bool_query.get(&world, a).is_err());
assert_eq!(i32_bool_query.get(&world, c).unwrap(), (&A(789), &B(1)));
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
assert!(world.despawn(a));
assert!(i32_query.get(&world, a).is_err());
}
#[test]
fn query_get_works_across_sparse_removal() {
// Regression test for: https://github.com/bevyengine/bevy/issues/6623
let mut world = World::new();
let a = world.spawn((TableStored("abc"), SparseStored(123))).id();
let b = world.spawn((TableStored("def"), SparseStored(456))).id();
let c = world
.spawn((TableStored("ghi"), SparseStored(789), B(1)))
.id();
let mut query = world.query::<&TableStored>();
assert_eq!(query.get(&world, a).unwrap(), &TableStored("abc"));
assert_eq!(query.get(&world, b).unwrap(), &TableStored("def"));
assert_eq!(query.get(&world, c).unwrap(), &TableStored("ghi"));
world.entity_mut(b).remove::<SparseStored>();
world.entity_mut(c).remove::<SparseStored>();
assert_eq!(query.get(&world, a).unwrap(), &TableStored("abc"));
assert_eq!(query.get(&world, b).unwrap(), &TableStored("def"));
assert_eq!(query.get(&world, c).unwrap(), &TableStored("ghi"));
}
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
#[test]
fn remove_tracking() {
let mut world = World::new();
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
let a = world.spawn((SparseStored(0), A(123))).id();
let b = world.spawn((SparseStored(1), A(123))).id();
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
world.entity_mut(a).despawn();
assert_eq!(
world.removed::<A>().collect::<Vec<_>>(),
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
&[a],
"despawning results in 'removed component' state for table components"
);
assert_eq!(
world.removed::<SparseStored>().collect::<Vec<_>>(),
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
&[a],
"despawning results in 'removed component' state for sparse set components"
);
world.entity_mut(b).insert(B(1));
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
assert_eq!(
world.removed::<A>().collect::<Vec<_>>(),
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
&[a],
"archetype moves does not result in 'removed component' state"
);
world.entity_mut(b).remove::<A>();
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
assert_eq!(
world.removed::<A>().collect::<Vec<_>>(),
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
&[a, b],
"removing a component results in a 'removed component' state"
);
world.clear_trackers();
assert_eq!(
world.removed::<A>().collect::<Vec<_>>(),
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
&[],
"clearing trackers clears removals"
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
);
assert_eq!(
world.removed::<SparseStored>().collect::<Vec<_>>(),
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
&[],
"clearing trackers clears removals"
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
);
assert_eq!(
world.removed::<B>().collect::<Vec<_>>(),
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
&[],
"clearing trackers clears removals"
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
);
// TODO: uncomment when world.clear() is implemented
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
// let c = world.spawn(("abc", 123)).id();
// let d = world.spawn(("abc", 123)).id();
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
// world.clear();
// assert_eq!(
// world.removed::<i32>(),
// &[c, d],
// "world clears result in 'removed component' states"
// );
// assert_eq!(
// world.removed::<&'static str>(),
// &[c, d, b],
// "world clears result in 'removed component' states"
// );
// assert_eq!(
// world.removed::<f64>(),
// &[b],
// "world clears result in 'removed component' states"
// );
}
#[test]
fn added_tracking() {
let mut world = World::new();
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
let a = world.spawn(A(123)).id();
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
assert_eq!(world.query::<&A>().iter(&world).count(), 1);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
assert_eq!(
world.query_filtered::<(), Added<A>>().iter(&world).count(),
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
1
);
assert_eq!(world.query::<&A>().iter(&world).count(), 1);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
assert_eq!(
world.query_filtered::<(), Added<A>>().iter(&world).count(),
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
1
);
assert!(world.query::<&A>().get(&world, a).is_ok());
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
assert!(world
.query_filtered::<(), Added<A>>()
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
.get(&world, a)
.is_ok());
assert!(world.query::<&A>().get(&world, a).is_ok());
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
assert!(world
.query_filtered::<(), Added<A>>()
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
.get(&world, a)
.is_ok());
world.clear_trackers();
assert_eq!(world.query::<&A>().iter(&world).count(), 1);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
assert_eq!(
world.query_filtered::<(), Added<A>>().iter(&world).count(),
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
0
);
assert_eq!(world.query::<&A>().iter(&world).count(), 1);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
assert_eq!(
world.query_filtered::<(), Added<A>>().iter(&world).count(),
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
0
);
assert!(world.query::<&A>().get(&world, a).is_ok());
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
assert!(world
.query_filtered::<(), Added<A>>()
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
.get(&world, a)
.is_err());
assert!(world.query::<&A>().get(&world, a).is_ok());
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
assert!(world
.query_filtered::<(), Added<A>>()
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
.get(&world, a)
.is_err());
}
#[test]
fn added_queries() {
let mut world = World::default();
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
let e1 = world.spawn(A(0)).id();
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
fn get_added<Com: Component>(world: &mut World) -> Vec<Entity> {
world
.query_filtered::<Entity, Added<Com>>()
.iter(world)
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
.collect::<Vec<Entity>>()
}
assert_eq!(get_added::<A>(&mut world), vec![e1]);
world.entity_mut(e1).insert(B(0));
assert_eq!(get_added::<A>(&mut world), vec![e1]);
assert_eq!(get_added::<B>(&mut world), vec![e1]);
world.clear_trackers();
assert!(get_added::<A>(&mut world).is_empty());
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
let e2 = world.spawn((A(1), B(1))).id();
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
assert_eq!(get_added::<A>(&mut world), vec![e2]);
assert_eq!(get_added::<B>(&mut world), vec![e2]);
let added = world
.query_filtered::<Entity, (Added<A>, Added<B>)>()
.iter(&world)
.collect::<Vec<Entity>>();
assert_eq!(added, vec![e2]);
}
#[test]
Reliable change detection (#1471) # Problem Definition The current change tracking (via flags for both components and resources) fails to detect changes made by systems that are scheduled to run earlier in the frame than they are. This issue is discussed at length in [#68](https://github.com/bevyengine/bevy/issues/68) and [#54](https://github.com/bevyengine/bevy/issues/54). This is very much a draft PR, and contributions are welcome and needed. # Criteria 1. Each change is detected at least once, no matter the ordering. 2. Each change is detected at most once, no matter the ordering. 3. Changes should be detected the same frame that they are made. 4. Competitive ergonomics. Ideally does not require opting-in. 5. Low CPU overhead of computation. 6. Memory efficient. This must not increase over time, except where the number of entities / resources does. 7. Changes should not be lost for systems that don't run. 8. A frame needs to act as a pure function. Given the same set of entities / components it needs to produce the same end state without side-effects. **Exact** change-tracking proposals satisfy criteria 1 and 2. **Conservative** change-tracking proposals satisfy criteria 1 but not 2. **Flaky** change tracking proposals satisfy criteria 2 but not 1. # Code Base Navigation There are three types of flags: - `Added`: A piece of data was added to an entity / `Resources`. - `Mutated`: A piece of data was able to be modified, because its `DerefMut` was accessed - `Changed`: The bitwise OR of `Added` and `Changed` The special behavior of `ChangedRes`, with respect to the scheduler is being removed in [#1313](https://github.com/bevyengine/bevy/pull/1313) and does not need to be reproduced. `ChangedRes` and friends can be found in "bevy_ecs/core/resources/resource_query.rs". The `Flags` trait for Components can be found in "bevy_ecs/core/query.rs". `ComponentFlags` are stored in "bevy_ecs/core/archetypes.rs", defined on line 446. # Proposals **Proposal 5 was selected for implementation.** ## Proposal 0: No Change Detection The baseline, where computations are performed on everything regardless of whether it changed. **Type:** Conservative **Pros:** - already implemented - will never miss events - no overhead **Cons:** - tons of repeated work - doesn't allow users to avoid repeating work (or monitoring for other changes) ## Proposal 1: Earlier-This-Tick Change Detection The current approach as of Bevy 0.4. Flags are set, and then flushed at the end of each frame. **Type:** Flaky **Pros:** - already implemented - simple to understand - low memory overhead (2 bits per component) - low time overhead (clear every flag once per frame) **Cons:** - misses systems based on ordering - systems that don't run every frame miss changes - duplicates detection when looping - can lead to unresolvable circular dependencies ## Proposal 2: Two-Tick Change Detection Flags persist for two frames, using a double-buffer system identical to that used in events. A change is observed if it is found in either the current frame's list of changes or the previous frame's. **Type:** Conservative **Pros:** - easy to understand - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - can result in a great deal of duplicated work - systems that don't run every frame miss changes - duplicates detection when looping ## Proposal 3: Last-Tick Change Detection Flags persist for two frames, using a double-buffer system identical to that used in events. A change is observed if it is found in the previous frame's list of changes. **Type:** Exact **Pros:** - exact - easy to understand - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - change detection is always delayed, possibly causing painful chained delays - systems that don't run every frame miss changes - duplicates detection when looping ## Proposal 4: Flag-Doubling Change Detection Combine Proposal 2 and Proposal 3. Differentiate between `JustChanged` (current behavior) and `Changed` (Proposal 3). Pack this data into the flags according to [this implementation proposal](https://github.com/bevyengine/bevy/issues/68#issuecomment-769174804). **Type:** Flaky + Exact **Pros:** - allows users to acc - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - users must specify the type of change detection required - still quite fragile to system ordering effects when using the flaky `JustChanged` form - cannot get immediate + exact results - systems that don't run every frame miss changes - duplicates detection when looping ## [SELECTED] Proposal 5: Generation-Counter Change Detection A global counter is increased after each system is run. Each component saves the time of last mutation, and each system saves the time of last execution. Mutation is detected when the component's counter is greater than the system's counter. Discussed [here](https://github.com/bevyengine/bevy/issues/68#issuecomment-769174804). How to handle addition detection is unsolved; the current proposal is to use the highest bit of the counter as in proposal 1. **Type:** Exact (for mutations), flaky (for additions) **Pros:** - low time overhead (set component counter on access, set system counter after execution) - robust to systems that don't run every frame - robust to systems that loop **Cons:** - moderately complex implementation - must be modified as systems are inserted dynamically - medium memory overhead (4 bytes per component + system) - unsolved addition detection ## Proposal 6: System-Data Change Detection For each system, track which system's changes it has seen. This approach is only worth fully designing and implementing if Proposal 5 fails in some way. **Type:** Exact **Pros:** - exact - conceptually simple **Cons:** - requires storing data on each system - implementation is complex - must be modified as systems are inserted dynamically ## Proposal 7: Total-Order Change Detection Discussed [here](https://github.com/bevyengine/bevy/issues/68#issuecomment-754326523). This proposal is somewhat complicated by the new scheduler, but I believe it should still be conceptually feasible. This approach is only worth fully designing and implementing if Proposal 5 fails in some way. **Type:** Exact **Pros:** - exact - efficient data storage relative to other exact proposals **Cons:** - requires access to the scheduler - complex implementation and difficulty grokking - must be modified as systems are inserted dynamically # Tests - We will need to verify properties 1, 2, 3, 7 and 8. Priority: 1 > 2 = 3 > 8 > 7 - Ideally we can use identical user-facing syntax for all proposals, allowing us to re-use the same syntax for each. - When writing tests, we need to carefully specify order using explicit dependencies. - These tests will need to be duplicated for both components and resources. - We need to be sure to handle cases where ambiguous system orders exist. `changing_system` is always the system that makes the changes, and `detecting_system` always detects the changes. The component / resource changed will be simple boolean wrapper structs. ## Basic Added / Mutated / Changed 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs before `detecting_system` - verify at the end of tick 2 ## At Least Once 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs after `detecting_system` - verify at the end of tick 2 ## At Most Once 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs once before `detecting_system` - increment a counter based on the number of changes detected - verify at the end of tick 2 ## Fast Detection 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs before `detecting_system` - verify at the end of tick 1 ## Ambiguous System Ordering Robustness 2 x 3 x 2 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs [before/after] `detecting_system` in tick 1 - `changing_system` runs [after/before] `detecting_system` in tick 2 ## System Pausing 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs in tick 1, then is disabled by run criteria - `detecting_system` is disabled by run criteria until it is run once during tick 3 - verify at the end of tick 3 ## Addition Causes Mutation 2 design: - Resources vs. Components - `adding_system_1` adds a component / resource - `adding system_2` adds the same component / resource - verify the `Mutated` flag at the end of the tick - verify the `Added` flag at the end of the tick First check tests for: https://github.com/bevyengine/bevy/issues/333 Second check tests for: https://github.com/bevyengine/bevy/issues/1443 ## Changes Made By Commands - `adding_system` runs in Update in tick 1, and sends a command to add a component - `detecting_system` runs in Update in tick 1 and 2, after `adding_system` - We can't detect the changes in tick 1, since they haven't been processed yet - If we were to track these changes as being emitted by `adding_system`, we can't detect the changes in tick 2 either, since `detecting_system` has already run once after `adding_system` :( # Benchmarks See: [general advice](https://github.com/bevyengine/bevy/blob/master/docs/profiling.md), [Criterion crate](https://github.com/bheisler/criterion.rs) There are several critical parameters to vary: 1. entity count (1 to 10^9) 2. fraction of entities that are changed (0% to 100%) 3. cost to perform work on changed entities, i.e. workload (1 ns to 1s) 1 and 2 should be varied between benchmark runs. 3 can be added on computationally. We want to measure: - memory cost - run time We should collect these measurements across several frames (100?) to reduce bootup effects and accurately measure the mean, variance and drift. Entity-component change detection is much more important to benchmark than resource change detection, due to the orders of magnitude higher number of pieces of data. No change detection at all should be included in benchmarks as a second control for cases where missing changes is unacceptable. ## Graphs 1. y: performance, x: log_10(entity count), color: proposal, facet: performance metric. Set cost to perform work to 0. 2. y: run time, x: cost to perform work, color: proposal, facet: fraction changed. Set number of entities to 10^6 3. y: memory, x: frames, color: proposal # Conclusions 1. Is the theoretical categorization of the proposals correct according to our tests? 2. How does the performance of the proposals compare without any load? 3. How does the performance of the proposals compare with realistic loads? 4. At what workload does more exact change tracking become worth the (presumably) higher overhead? 5. When does adding change-detection to save on work become worthwhile? 6. Is there enough divergence in performance between the best solutions in each class to ship more than one change-tracking solution? # Implementation Plan 1. Write a test suite. 2. Verify that tests fail for existing approach. 3. Write a benchmark suite. 4. Get performance numbers for existing approach. 5. Implement, test and benchmark various solutions using a Git branch per proposal. 6. Create a draft PR with all solutions and present results to team. 7. Select a solution and replace existing change detection. Co-authored-by: Brice DAVIER <bricedavier@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2021-03-19 17:53:26 +00:00
fn changed_trackers() {
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
let mut world = World::default();
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
let e1 = world.spawn((A(0), B(0))).id();
let e2 = world.spawn((A(0), B(0))).id();
let e3 = world.spawn((A(0), B(0))).id();
world.spawn((A(0), B(0)));
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
Reliable change detection (#1471) # Problem Definition The current change tracking (via flags for both components and resources) fails to detect changes made by systems that are scheduled to run earlier in the frame than they are. This issue is discussed at length in [#68](https://github.com/bevyengine/bevy/issues/68) and [#54](https://github.com/bevyengine/bevy/issues/54). This is very much a draft PR, and contributions are welcome and needed. # Criteria 1. Each change is detected at least once, no matter the ordering. 2. Each change is detected at most once, no matter the ordering. 3. Changes should be detected the same frame that they are made. 4. Competitive ergonomics. Ideally does not require opting-in. 5. Low CPU overhead of computation. 6. Memory efficient. This must not increase over time, except where the number of entities / resources does. 7. Changes should not be lost for systems that don't run. 8. A frame needs to act as a pure function. Given the same set of entities / components it needs to produce the same end state without side-effects. **Exact** change-tracking proposals satisfy criteria 1 and 2. **Conservative** change-tracking proposals satisfy criteria 1 but not 2. **Flaky** change tracking proposals satisfy criteria 2 but not 1. # Code Base Navigation There are three types of flags: - `Added`: A piece of data was added to an entity / `Resources`. - `Mutated`: A piece of data was able to be modified, because its `DerefMut` was accessed - `Changed`: The bitwise OR of `Added` and `Changed` The special behavior of `ChangedRes`, with respect to the scheduler is being removed in [#1313](https://github.com/bevyengine/bevy/pull/1313) and does not need to be reproduced. `ChangedRes` and friends can be found in "bevy_ecs/core/resources/resource_query.rs". The `Flags` trait for Components can be found in "bevy_ecs/core/query.rs". `ComponentFlags` are stored in "bevy_ecs/core/archetypes.rs", defined on line 446. # Proposals **Proposal 5 was selected for implementation.** ## Proposal 0: No Change Detection The baseline, where computations are performed on everything regardless of whether it changed. **Type:** Conservative **Pros:** - already implemented - will never miss events - no overhead **Cons:** - tons of repeated work - doesn't allow users to avoid repeating work (or monitoring for other changes) ## Proposal 1: Earlier-This-Tick Change Detection The current approach as of Bevy 0.4. Flags are set, and then flushed at the end of each frame. **Type:** Flaky **Pros:** - already implemented - simple to understand - low memory overhead (2 bits per component) - low time overhead (clear every flag once per frame) **Cons:** - misses systems based on ordering - systems that don't run every frame miss changes - duplicates detection when looping - can lead to unresolvable circular dependencies ## Proposal 2: Two-Tick Change Detection Flags persist for two frames, using a double-buffer system identical to that used in events. A change is observed if it is found in either the current frame's list of changes or the previous frame's. **Type:** Conservative **Pros:** - easy to understand - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - can result in a great deal of duplicated work - systems that don't run every frame miss changes - duplicates detection when looping ## Proposal 3: Last-Tick Change Detection Flags persist for two frames, using a double-buffer system identical to that used in events. A change is observed if it is found in the previous frame's list of changes. **Type:** Exact **Pros:** - exact - easy to understand - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - change detection is always delayed, possibly causing painful chained delays - systems that don't run every frame miss changes - duplicates detection when looping ## Proposal 4: Flag-Doubling Change Detection Combine Proposal 2 and Proposal 3. Differentiate between `JustChanged` (current behavior) and `Changed` (Proposal 3). Pack this data into the flags according to [this implementation proposal](https://github.com/bevyengine/bevy/issues/68#issuecomment-769174804). **Type:** Flaky + Exact **Pros:** - allows users to acc - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - users must specify the type of change detection required - still quite fragile to system ordering effects when using the flaky `JustChanged` form - cannot get immediate + exact results - systems that don't run every frame miss changes - duplicates detection when looping ## [SELECTED] Proposal 5: Generation-Counter Change Detection A global counter is increased after each system is run. Each component saves the time of last mutation, and each system saves the time of last execution. Mutation is detected when the component's counter is greater than the system's counter. Discussed [here](https://github.com/bevyengine/bevy/issues/68#issuecomment-769174804). How to handle addition detection is unsolved; the current proposal is to use the highest bit of the counter as in proposal 1. **Type:** Exact (for mutations), flaky (for additions) **Pros:** - low time overhead (set component counter on access, set system counter after execution) - robust to systems that don't run every frame - robust to systems that loop **Cons:** - moderately complex implementation - must be modified as systems are inserted dynamically - medium memory overhead (4 bytes per component + system) - unsolved addition detection ## Proposal 6: System-Data Change Detection For each system, track which system's changes it has seen. This approach is only worth fully designing and implementing if Proposal 5 fails in some way. **Type:** Exact **Pros:** - exact - conceptually simple **Cons:** - requires storing data on each system - implementation is complex - must be modified as systems are inserted dynamically ## Proposal 7: Total-Order Change Detection Discussed [here](https://github.com/bevyengine/bevy/issues/68#issuecomment-754326523). This proposal is somewhat complicated by the new scheduler, but I believe it should still be conceptually feasible. This approach is only worth fully designing and implementing if Proposal 5 fails in some way. **Type:** Exact **Pros:** - exact - efficient data storage relative to other exact proposals **Cons:** - requires access to the scheduler - complex implementation and difficulty grokking - must be modified as systems are inserted dynamically # Tests - We will need to verify properties 1, 2, 3, 7 and 8. Priority: 1 > 2 = 3 > 8 > 7 - Ideally we can use identical user-facing syntax for all proposals, allowing us to re-use the same syntax for each. - When writing tests, we need to carefully specify order using explicit dependencies. - These tests will need to be duplicated for both components and resources. - We need to be sure to handle cases where ambiguous system orders exist. `changing_system` is always the system that makes the changes, and `detecting_system` always detects the changes. The component / resource changed will be simple boolean wrapper structs. ## Basic Added / Mutated / Changed 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs before `detecting_system` - verify at the end of tick 2 ## At Least Once 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs after `detecting_system` - verify at the end of tick 2 ## At Most Once 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs once before `detecting_system` - increment a counter based on the number of changes detected - verify at the end of tick 2 ## Fast Detection 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs before `detecting_system` - verify at the end of tick 1 ## Ambiguous System Ordering Robustness 2 x 3 x 2 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs [before/after] `detecting_system` in tick 1 - `changing_system` runs [after/before] `detecting_system` in tick 2 ## System Pausing 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs in tick 1, then is disabled by run criteria - `detecting_system` is disabled by run criteria until it is run once during tick 3 - verify at the end of tick 3 ## Addition Causes Mutation 2 design: - Resources vs. Components - `adding_system_1` adds a component / resource - `adding system_2` adds the same component / resource - verify the `Mutated` flag at the end of the tick - verify the `Added` flag at the end of the tick First check tests for: https://github.com/bevyengine/bevy/issues/333 Second check tests for: https://github.com/bevyengine/bevy/issues/1443 ## Changes Made By Commands - `adding_system` runs in Update in tick 1, and sends a command to add a component - `detecting_system` runs in Update in tick 1 and 2, after `adding_system` - We can't detect the changes in tick 1, since they haven't been processed yet - If we were to track these changes as being emitted by `adding_system`, we can't detect the changes in tick 2 either, since `detecting_system` has already run once after `adding_system` :( # Benchmarks See: [general advice](https://github.com/bevyengine/bevy/blob/master/docs/profiling.md), [Criterion crate](https://github.com/bheisler/criterion.rs) There are several critical parameters to vary: 1. entity count (1 to 10^9) 2. fraction of entities that are changed (0% to 100%) 3. cost to perform work on changed entities, i.e. workload (1 ns to 1s) 1 and 2 should be varied between benchmark runs. 3 can be added on computationally. We want to measure: - memory cost - run time We should collect these measurements across several frames (100?) to reduce bootup effects and accurately measure the mean, variance and drift. Entity-component change detection is much more important to benchmark than resource change detection, due to the orders of magnitude higher number of pieces of data. No change detection at all should be included in benchmarks as a second control for cases where missing changes is unacceptable. ## Graphs 1. y: performance, x: log_10(entity count), color: proposal, facet: performance metric. Set cost to perform work to 0. 2. y: run time, x: cost to perform work, color: proposal, facet: fraction changed. Set number of entities to 10^6 3. y: memory, x: frames, color: proposal # Conclusions 1. Is the theoretical categorization of the proposals correct according to our tests? 2. How does the performance of the proposals compare without any load? 3. How does the performance of the proposals compare with realistic loads? 4. At what workload does more exact change tracking become worth the (presumably) higher overhead? 5. When does adding change-detection to save on work become worthwhile? 6. Is there enough divergence in performance between the best solutions in each class to ship more than one change-tracking solution? # Implementation Plan 1. Write a test suite. 2. Verify that tests fail for existing approach. 3. Write a benchmark suite. 4. Get performance numbers for existing approach. 5. Implement, test and benchmark various solutions using a Git branch per proposal. 6. Create a draft PR with all solutions and present results to team. 7. Select a solution and replace existing change detection. Co-authored-by: Brice DAVIER <bricedavier@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2021-03-19 17:53:26 +00:00
world.clear_trackers();
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
for (i, mut a) in world.query::<&mut A>().iter_mut(&mut world).enumerate() {
if i % 2 == 0 {
a.0 += 1;
}
}
Rename `WorldQueryData` & `WorldQueryFilter` to `QueryData` & `QueryFilter` (#10779) # Rename `WorldQueryData` & `WorldQueryFilter` to `QueryData` & `QueryFilter` Fixes #10776 ## Solution Traits `WorldQueryData` & `WorldQueryFilter` were renamed to `QueryData` and `QueryFilter`, respectively. Related Trait types were also renamed. --- ## Changelog - Trait `WorldQueryData` has been renamed to `QueryData`. Derive macro's `QueryData` attribute `world_query_data` has been renamed to `query_data`. - Trait `WorldQueryFilter` has been renamed to `QueryFilter`. Derive macro's `QueryFilter` attribute `world_query_filter` has been renamed to `query_filter`. - Trait's `ExtractComponent` type `Query` has been renamed to `Data`. - Trait's `GetBatchData` types `Query` & `QueryFilter` has been renamed to `Data` & `Filter`, respectively. - Trait's `ExtractInstance` type `Query` has been renamed to `Data`. - Trait's `ViewNode` type `ViewQuery` has been renamed to `ViewData`. - Trait's `RenderCommand` types `ViewWorldQuery` & `ItemWorldQuery` has been renamed to `ViewData` & `ItemData`, respectively. ## Migration Guide Note: if merged before 0.13 is released, this should instead modify the migration guide of #10776 with the updated names. - Rename `WorldQueryData` & `WorldQueryFilter` trait usages to `QueryData` & `QueryFilter` and their respective derive macro attributes `world_query_data` & `world_query_filter` to `query_data` & `query_filter`. - Rename the following trait type usages: - Trait's `ExtractComponent` type `Query` to `Data`. - Trait's `GetBatchData` type `Query` to `Data`. - Trait's `ExtractInstance` type `Query` to `Data`. - Trait's `ViewNode` type `ViewQuery` to `ViewData`' - Trait's `RenderCommand` types `ViewWolrdQuery` & `ItemWorldQuery` to `ViewData` & `ItemData`, respectively. ```rust // Before #[derive(WorldQueryData)] #[world_query_data(derive(Debug))] struct EmptyQuery { empty: (), } // After #[derive(QueryData)] #[query_data(derive(Debug))] struct EmptyQuery { empty: (), } // Before #[derive(WorldQueryFilter)] struct CustomQueryFilter<T: Component, P: Component> { _c: With<ComponentC>, _d: With<ComponentD>, _or: Or<(Added<ComponentC>, Changed<ComponentD>, Without<ComponentZ>)>, _generic_tuple: (With<T>, With<P>), } // After #[derive(QueryFilter)] struct CustomQueryFilter<T: Component, P: Component> { _c: With<ComponentC>, _d: With<ComponentD>, _or: Or<(Added<ComponentC>, Changed<ComponentD>, Without<ComponentZ>)>, _generic_tuple: (With<T>, With<P>), } // Before impl ExtractComponent for ContrastAdaptiveSharpeningSettings { type Query = &'static Self; type Filter = With<Camera>; type Out = (DenoiseCAS, CASUniform); fn extract_component(item: QueryItem<Self::Query>) -> Option<Self::Out> { //... } } // After impl ExtractComponent for ContrastAdaptiveSharpeningSettings { type Data = &'static Self; type Filter = With<Camera>; type Out = (DenoiseCAS, CASUniform); fn extract_component(item: QueryItem<Self::Data>) -> Option<Self::Out> { //... } } // Before impl GetBatchData for MeshPipeline { type Param = SRes<RenderMeshInstances>; type Query = Entity; type QueryFilter = With<Mesh3d>; type CompareData = (MaterialBindGroupId, AssetId<Mesh>); type BufferData = MeshUniform; fn get_batch_data( mesh_instances: &SystemParamItem<Self::Param>, entity: &QueryItem<Self::Query>, ) -> (Self::BufferData, Option<Self::CompareData>) { // .... } } // After impl GetBatchData for MeshPipeline { type Param = SRes<RenderMeshInstances>; type Data = Entity; type Filter = With<Mesh3d>; type CompareData = (MaterialBindGroupId, AssetId<Mesh>); type BufferData = MeshUniform; fn get_batch_data( mesh_instances: &SystemParamItem<Self::Param>, entity: &QueryItem<Self::Data>, ) -> (Self::BufferData, Option<Self::CompareData>) { // .... } } // Before impl<A> ExtractInstance for AssetId<A> where A: Asset, { type Query = Read<Handle<A>>; type Filter = (); fn extract(item: QueryItem<'_, Self::Query>) -> Option<Self> { Some(item.id()) } } // After impl<A> ExtractInstance for AssetId<A> where A: Asset, { type Data = Read<Handle<A>>; type Filter = (); fn extract(item: QueryItem<'_, Self::Data>) -> Option<Self> { Some(item.id()) } } // Before impl ViewNode for PostProcessNode { type ViewQuery = ( &'static ViewTarget, &'static PostProcessSettings, ); fn run( &self, _graph: &mut RenderGraphContext, render_context: &mut RenderContext, (view_target, _post_process_settings): QueryItem<Self::ViewQuery>, world: &World, ) -> Result<(), NodeRunError> { // ... } } // After impl ViewNode for PostProcessNode { type ViewData = ( &'static ViewTarget, &'static PostProcessSettings, ); fn run( &self, _graph: &mut RenderGraphContext, render_context: &mut RenderContext, (view_target, _post_process_settings): QueryItem<Self::ViewData>, world: &World, ) -> Result<(), NodeRunError> { // ... } } // Before impl<P: CachedRenderPipelinePhaseItem> RenderCommand<P> for SetItemPipeline { type Param = SRes<PipelineCache>; type ViewWorldQuery = (); type ItemWorldQuery = (); #[inline] fn render<'w>( item: &P, _view: (), _entity: (), pipeline_cache: SystemParamItem<'w, '_, Self::Param>, pass: &mut TrackedRenderPass<'w>, ) -> RenderCommandResult { // ... } } // After impl<P: CachedRenderPipelinePhaseItem> RenderCommand<P> for SetItemPipeline { type Param = SRes<PipelineCache>; type ViewData = (); type ItemData = (); #[inline] fn render<'w>( item: &P, _view: (), _entity: (), pipeline_cache: SystemParamItem<'w, '_, Self::Param>, pass: &mut TrackedRenderPass<'w>, ) -> RenderCommandResult { // ... } } ```
2023-12-12 19:45:50 +00:00
fn get_filtered<F: QueryFilter>(world: &mut World) -> Vec<Entity> {
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
world
.query_filtered::<Entity, F>()
.iter(world)
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
.collect::<Vec<Entity>>()
}
Reliable change detection (#1471) # Problem Definition The current change tracking (via flags for both components and resources) fails to detect changes made by systems that are scheduled to run earlier in the frame than they are. This issue is discussed at length in [#68](https://github.com/bevyengine/bevy/issues/68) and [#54](https://github.com/bevyengine/bevy/issues/54). This is very much a draft PR, and contributions are welcome and needed. # Criteria 1. Each change is detected at least once, no matter the ordering. 2. Each change is detected at most once, no matter the ordering. 3. Changes should be detected the same frame that they are made. 4. Competitive ergonomics. Ideally does not require opting-in. 5. Low CPU overhead of computation. 6. Memory efficient. This must not increase over time, except where the number of entities / resources does. 7. Changes should not be lost for systems that don't run. 8. A frame needs to act as a pure function. Given the same set of entities / components it needs to produce the same end state without side-effects. **Exact** change-tracking proposals satisfy criteria 1 and 2. **Conservative** change-tracking proposals satisfy criteria 1 but not 2. **Flaky** change tracking proposals satisfy criteria 2 but not 1. # Code Base Navigation There are three types of flags: - `Added`: A piece of data was added to an entity / `Resources`. - `Mutated`: A piece of data was able to be modified, because its `DerefMut` was accessed - `Changed`: The bitwise OR of `Added` and `Changed` The special behavior of `ChangedRes`, with respect to the scheduler is being removed in [#1313](https://github.com/bevyengine/bevy/pull/1313) and does not need to be reproduced. `ChangedRes` and friends can be found in "bevy_ecs/core/resources/resource_query.rs". The `Flags` trait for Components can be found in "bevy_ecs/core/query.rs". `ComponentFlags` are stored in "bevy_ecs/core/archetypes.rs", defined on line 446. # Proposals **Proposal 5 was selected for implementation.** ## Proposal 0: No Change Detection The baseline, where computations are performed on everything regardless of whether it changed. **Type:** Conservative **Pros:** - already implemented - will never miss events - no overhead **Cons:** - tons of repeated work - doesn't allow users to avoid repeating work (or monitoring for other changes) ## Proposal 1: Earlier-This-Tick Change Detection The current approach as of Bevy 0.4. Flags are set, and then flushed at the end of each frame. **Type:** Flaky **Pros:** - already implemented - simple to understand - low memory overhead (2 bits per component) - low time overhead (clear every flag once per frame) **Cons:** - misses systems based on ordering - systems that don't run every frame miss changes - duplicates detection when looping - can lead to unresolvable circular dependencies ## Proposal 2: Two-Tick Change Detection Flags persist for two frames, using a double-buffer system identical to that used in events. A change is observed if it is found in either the current frame's list of changes or the previous frame's. **Type:** Conservative **Pros:** - easy to understand - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - can result in a great deal of duplicated work - systems that don't run every frame miss changes - duplicates detection when looping ## Proposal 3: Last-Tick Change Detection Flags persist for two frames, using a double-buffer system identical to that used in events. A change is observed if it is found in the previous frame's list of changes. **Type:** Exact **Pros:** - exact - easy to understand - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - change detection is always delayed, possibly causing painful chained delays - systems that don't run every frame miss changes - duplicates detection when looping ## Proposal 4: Flag-Doubling Change Detection Combine Proposal 2 and Proposal 3. Differentiate between `JustChanged` (current behavior) and `Changed` (Proposal 3). Pack this data into the flags according to [this implementation proposal](https://github.com/bevyengine/bevy/issues/68#issuecomment-769174804). **Type:** Flaky + Exact **Pros:** - allows users to acc - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - users must specify the type of change detection required - still quite fragile to system ordering effects when using the flaky `JustChanged` form - cannot get immediate + exact results - systems that don't run every frame miss changes - duplicates detection when looping ## [SELECTED] Proposal 5: Generation-Counter Change Detection A global counter is increased after each system is run. Each component saves the time of last mutation, and each system saves the time of last execution. Mutation is detected when the component's counter is greater than the system's counter. Discussed [here](https://github.com/bevyengine/bevy/issues/68#issuecomment-769174804). How to handle addition detection is unsolved; the current proposal is to use the highest bit of the counter as in proposal 1. **Type:** Exact (for mutations), flaky (for additions) **Pros:** - low time overhead (set component counter on access, set system counter after execution) - robust to systems that don't run every frame - robust to systems that loop **Cons:** - moderately complex implementation - must be modified as systems are inserted dynamically - medium memory overhead (4 bytes per component + system) - unsolved addition detection ## Proposal 6: System-Data Change Detection For each system, track which system's changes it has seen. This approach is only worth fully designing and implementing if Proposal 5 fails in some way. **Type:** Exact **Pros:** - exact - conceptually simple **Cons:** - requires storing data on each system - implementation is complex - must be modified as systems are inserted dynamically ## Proposal 7: Total-Order Change Detection Discussed [here](https://github.com/bevyengine/bevy/issues/68#issuecomment-754326523). This proposal is somewhat complicated by the new scheduler, but I believe it should still be conceptually feasible. This approach is only worth fully designing and implementing if Proposal 5 fails in some way. **Type:** Exact **Pros:** - exact - efficient data storage relative to other exact proposals **Cons:** - requires access to the scheduler - complex implementation and difficulty grokking - must be modified as systems are inserted dynamically # Tests - We will need to verify properties 1, 2, 3, 7 and 8. Priority: 1 > 2 = 3 > 8 > 7 - Ideally we can use identical user-facing syntax for all proposals, allowing us to re-use the same syntax for each. - When writing tests, we need to carefully specify order using explicit dependencies. - These tests will need to be duplicated for both components and resources. - We need to be sure to handle cases where ambiguous system orders exist. `changing_system` is always the system that makes the changes, and `detecting_system` always detects the changes. The component / resource changed will be simple boolean wrapper structs. ## Basic Added / Mutated / Changed 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs before `detecting_system` - verify at the end of tick 2 ## At Least Once 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs after `detecting_system` - verify at the end of tick 2 ## At Most Once 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs once before `detecting_system` - increment a counter based on the number of changes detected - verify at the end of tick 2 ## Fast Detection 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs before `detecting_system` - verify at the end of tick 1 ## Ambiguous System Ordering Robustness 2 x 3 x 2 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs [before/after] `detecting_system` in tick 1 - `changing_system` runs [after/before] `detecting_system` in tick 2 ## System Pausing 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs in tick 1, then is disabled by run criteria - `detecting_system` is disabled by run criteria until it is run once during tick 3 - verify at the end of tick 3 ## Addition Causes Mutation 2 design: - Resources vs. Components - `adding_system_1` adds a component / resource - `adding system_2` adds the same component / resource - verify the `Mutated` flag at the end of the tick - verify the `Added` flag at the end of the tick First check tests for: https://github.com/bevyengine/bevy/issues/333 Second check tests for: https://github.com/bevyengine/bevy/issues/1443 ## Changes Made By Commands - `adding_system` runs in Update in tick 1, and sends a command to add a component - `detecting_system` runs in Update in tick 1 and 2, after `adding_system` - We can't detect the changes in tick 1, since they haven't been processed yet - If we were to track these changes as being emitted by `adding_system`, we can't detect the changes in tick 2 either, since `detecting_system` has already run once after `adding_system` :( # Benchmarks See: [general advice](https://github.com/bevyengine/bevy/blob/master/docs/profiling.md), [Criterion crate](https://github.com/bheisler/criterion.rs) There are several critical parameters to vary: 1. entity count (1 to 10^9) 2. fraction of entities that are changed (0% to 100%) 3. cost to perform work on changed entities, i.e. workload (1 ns to 1s) 1 and 2 should be varied between benchmark runs. 3 can be added on computationally. We want to measure: - memory cost - run time We should collect these measurements across several frames (100?) to reduce bootup effects and accurately measure the mean, variance and drift. Entity-component change detection is much more important to benchmark than resource change detection, due to the orders of magnitude higher number of pieces of data. No change detection at all should be included in benchmarks as a second control for cases where missing changes is unacceptable. ## Graphs 1. y: performance, x: log_10(entity count), color: proposal, facet: performance metric. Set cost to perform work to 0. 2. y: run time, x: cost to perform work, color: proposal, facet: fraction changed. Set number of entities to 10^6 3. y: memory, x: frames, color: proposal # Conclusions 1. Is the theoretical categorization of the proposals correct according to our tests? 2. How does the performance of the proposals compare without any load? 3. How does the performance of the proposals compare with realistic loads? 4. At what workload does more exact change tracking become worth the (presumably) higher overhead? 5. When does adding change-detection to save on work become worthwhile? 6. Is there enough divergence in performance between the best solutions in each class to ship more than one change-tracking solution? # Implementation Plan 1. Write a test suite. 2. Verify that tests fail for existing approach. 3. Write a benchmark suite. 4. Get performance numbers for existing approach. 5. Implement, test and benchmark various solutions using a Git branch per proposal. 6. Create a draft PR with all solutions and present results to team. 7. Select a solution and replace existing change detection. Co-authored-by: Brice DAVIER <bricedavier@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2021-03-19 17:53:26 +00:00
assert_eq!(get_filtered::<Changed<A>>(&mut world), vec![e1, e3]);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
Reliable change detection (#1471) # Problem Definition The current change tracking (via flags for both components and resources) fails to detect changes made by systems that are scheduled to run earlier in the frame than they are. This issue is discussed at length in [#68](https://github.com/bevyengine/bevy/issues/68) and [#54](https://github.com/bevyengine/bevy/issues/54). This is very much a draft PR, and contributions are welcome and needed. # Criteria 1. Each change is detected at least once, no matter the ordering. 2. Each change is detected at most once, no matter the ordering. 3. Changes should be detected the same frame that they are made. 4. Competitive ergonomics. Ideally does not require opting-in. 5. Low CPU overhead of computation. 6. Memory efficient. This must not increase over time, except where the number of entities / resources does. 7. Changes should not be lost for systems that don't run. 8. A frame needs to act as a pure function. Given the same set of entities / components it needs to produce the same end state without side-effects. **Exact** change-tracking proposals satisfy criteria 1 and 2. **Conservative** change-tracking proposals satisfy criteria 1 but not 2. **Flaky** change tracking proposals satisfy criteria 2 but not 1. # Code Base Navigation There are three types of flags: - `Added`: A piece of data was added to an entity / `Resources`. - `Mutated`: A piece of data was able to be modified, because its `DerefMut` was accessed - `Changed`: The bitwise OR of `Added` and `Changed` The special behavior of `ChangedRes`, with respect to the scheduler is being removed in [#1313](https://github.com/bevyengine/bevy/pull/1313) and does not need to be reproduced. `ChangedRes` and friends can be found in "bevy_ecs/core/resources/resource_query.rs". The `Flags` trait for Components can be found in "bevy_ecs/core/query.rs". `ComponentFlags` are stored in "bevy_ecs/core/archetypes.rs", defined on line 446. # Proposals **Proposal 5 was selected for implementation.** ## Proposal 0: No Change Detection The baseline, where computations are performed on everything regardless of whether it changed. **Type:** Conservative **Pros:** - already implemented - will never miss events - no overhead **Cons:** - tons of repeated work - doesn't allow users to avoid repeating work (or monitoring for other changes) ## Proposal 1: Earlier-This-Tick Change Detection The current approach as of Bevy 0.4. Flags are set, and then flushed at the end of each frame. **Type:** Flaky **Pros:** - already implemented - simple to understand - low memory overhead (2 bits per component) - low time overhead (clear every flag once per frame) **Cons:** - misses systems based on ordering - systems that don't run every frame miss changes - duplicates detection when looping - can lead to unresolvable circular dependencies ## Proposal 2: Two-Tick Change Detection Flags persist for two frames, using a double-buffer system identical to that used in events. A change is observed if it is found in either the current frame's list of changes or the previous frame's. **Type:** Conservative **Pros:** - easy to understand - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - can result in a great deal of duplicated work - systems that don't run every frame miss changes - duplicates detection when looping ## Proposal 3: Last-Tick Change Detection Flags persist for two frames, using a double-buffer system identical to that used in events. A change is observed if it is found in the previous frame's list of changes. **Type:** Exact **Pros:** - exact - easy to understand - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - change detection is always delayed, possibly causing painful chained delays - systems that don't run every frame miss changes - duplicates detection when looping ## Proposal 4: Flag-Doubling Change Detection Combine Proposal 2 and Proposal 3. Differentiate between `JustChanged` (current behavior) and `Changed` (Proposal 3). Pack this data into the flags according to [this implementation proposal](https://github.com/bevyengine/bevy/issues/68#issuecomment-769174804). **Type:** Flaky + Exact **Pros:** - allows users to acc - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - users must specify the type of change detection required - still quite fragile to system ordering effects when using the flaky `JustChanged` form - cannot get immediate + exact results - systems that don't run every frame miss changes - duplicates detection when looping ## [SELECTED] Proposal 5: Generation-Counter Change Detection A global counter is increased after each system is run. Each component saves the time of last mutation, and each system saves the time of last execution. Mutation is detected when the component's counter is greater than the system's counter. Discussed [here](https://github.com/bevyengine/bevy/issues/68#issuecomment-769174804). How to handle addition detection is unsolved; the current proposal is to use the highest bit of the counter as in proposal 1. **Type:** Exact (for mutations), flaky (for additions) **Pros:** - low time overhead (set component counter on access, set system counter after execution) - robust to systems that don't run every frame - robust to systems that loop **Cons:** - moderately complex implementation - must be modified as systems are inserted dynamically - medium memory overhead (4 bytes per component + system) - unsolved addition detection ## Proposal 6: System-Data Change Detection For each system, track which system's changes it has seen. This approach is only worth fully designing and implementing if Proposal 5 fails in some way. **Type:** Exact **Pros:** - exact - conceptually simple **Cons:** - requires storing data on each system - implementation is complex - must be modified as systems are inserted dynamically ## Proposal 7: Total-Order Change Detection Discussed [here](https://github.com/bevyengine/bevy/issues/68#issuecomment-754326523). This proposal is somewhat complicated by the new scheduler, but I believe it should still be conceptually feasible. This approach is only worth fully designing and implementing if Proposal 5 fails in some way. **Type:** Exact **Pros:** - exact - efficient data storage relative to other exact proposals **Cons:** - requires access to the scheduler - complex implementation and difficulty grokking - must be modified as systems are inserted dynamically # Tests - We will need to verify properties 1, 2, 3, 7 and 8. Priority: 1 > 2 = 3 > 8 > 7 - Ideally we can use identical user-facing syntax for all proposals, allowing us to re-use the same syntax for each. - When writing tests, we need to carefully specify order using explicit dependencies. - These tests will need to be duplicated for both components and resources. - We need to be sure to handle cases where ambiguous system orders exist. `changing_system` is always the system that makes the changes, and `detecting_system` always detects the changes. The component / resource changed will be simple boolean wrapper structs. ## Basic Added / Mutated / Changed 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs before `detecting_system` - verify at the end of tick 2 ## At Least Once 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs after `detecting_system` - verify at the end of tick 2 ## At Most Once 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs once before `detecting_system` - increment a counter based on the number of changes detected - verify at the end of tick 2 ## Fast Detection 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs before `detecting_system` - verify at the end of tick 1 ## Ambiguous System Ordering Robustness 2 x 3 x 2 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs [before/after] `detecting_system` in tick 1 - `changing_system` runs [after/before] `detecting_system` in tick 2 ## System Pausing 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs in tick 1, then is disabled by run criteria - `detecting_system` is disabled by run criteria until it is run once during tick 3 - verify at the end of tick 3 ## Addition Causes Mutation 2 design: - Resources vs. Components - `adding_system_1` adds a component / resource - `adding system_2` adds the same component / resource - verify the `Mutated` flag at the end of the tick - verify the `Added` flag at the end of the tick First check tests for: https://github.com/bevyengine/bevy/issues/333 Second check tests for: https://github.com/bevyengine/bevy/issues/1443 ## Changes Made By Commands - `adding_system` runs in Update in tick 1, and sends a command to add a component - `detecting_system` runs in Update in tick 1 and 2, after `adding_system` - We can't detect the changes in tick 1, since they haven't been processed yet - If we were to track these changes as being emitted by `adding_system`, we can't detect the changes in tick 2 either, since `detecting_system` has already run once after `adding_system` :( # Benchmarks See: [general advice](https://github.com/bevyengine/bevy/blob/master/docs/profiling.md), [Criterion crate](https://github.com/bheisler/criterion.rs) There are several critical parameters to vary: 1. entity count (1 to 10^9) 2. fraction of entities that are changed (0% to 100%) 3. cost to perform work on changed entities, i.e. workload (1 ns to 1s) 1 and 2 should be varied between benchmark runs. 3 can be added on computationally. We want to measure: - memory cost - run time We should collect these measurements across several frames (100?) to reduce bootup effects and accurately measure the mean, variance and drift. Entity-component change detection is much more important to benchmark than resource change detection, due to the orders of magnitude higher number of pieces of data. No change detection at all should be included in benchmarks as a second control for cases where missing changes is unacceptable. ## Graphs 1. y: performance, x: log_10(entity count), color: proposal, facet: performance metric. Set cost to perform work to 0. 2. y: run time, x: cost to perform work, color: proposal, facet: fraction changed. Set number of entities to 10^6 3. y: memory, x: frames, color: proposal # Conclusions 1. Is the theoretical categorization of the proposals correct according to our tests? 2. How does the performance of the proposals compare without any load? 3. How does the performance of the proposals compare with realistic loads? 4. At what workload does more exact change tracking become worth the (presumably) higher overhead? 5. When does adding change-detection to save on work become worthwhile? 6. Is there enough divergence in performance between the best solutions in each class to ship more than one change-tracking solution? # Implementation Plan 1. Write a test suite. 2. Verify that tests fail for existing approach. 3. Write a benchmark suite. 4. Get performance numbers for existing approach. 5. Implement, test and benchmark various solutions using a Git branch per proposal. 6. Create a draft PR with all solutions and present results to team. 7. Select a solution and replace existing change detection. Co-authored-by: Brice DAVIER <bricedavier@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2021-03-19 17:53:26 +00:00
// ensure changing an entity's archetypes also moves its changed state
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
world.entity_mut(e1).insert(C);
Reliable change detection (#1471) # Problem Definition The current change tracking (via flags for both components and resources) fails to detect changes made by systems that are scheduled to run earlier in the frame than they are. This issue is discussed at length in [#68](https://github.com/bevyengine/bevy/issues/68) and [#54](https://github.com/bevyengine/bevy/issues/54). This is very much a draft PR, and contributions are welcome and needed. # Criteria 1. Each change is detected at least once, no matter the ordering. 2. Each change is detected at most once, no matter the ordering. 3. Changes should be detected the same frame that they are made. 4. Competitive ergonomics. Ideally does not require opting-in. 5. Low CPU overhead of computation. 6. Memory efficient. This must not increase over time, except where the number of entities / resources does. 7. Changes should not be lost for systems that don't run. 8. A frame needs to act as a pure function. Given the same set of entities / components it needs to produce the same end state without side-effects. **Exact** change-tracking proposals satisfy criteria 1 and 2. **Conservative** change-tracking proposals satisfy criteria 1 but not 2. **Flaky** change tracking proposals satisfy criteria 2 but not 1. # Code Base Navigation There are three types of flags: - `Added`: A piece of data was added to an entity / `Resources`. - `Mutated`: A piece of data was able to be modified, because its `DerefMut` was accessed - `Changed`: The bitwise OR of `Added` and `Changed` The special behavior of `ChangedRes`, with respect to the scheduler is being removed in [#1313](https://github.com/bevyengine/bevy/pull/1313) and does not need to be reproduced. `ChangedRes` and friends can be found in "bevy_ecs/core/resources/resource_query.rs". The `Flags` trait for Components can be found in "bevy_ecs/core/query.rs". `ComponentFlags` are stored in "bevy_ecs/core/archetypes.rs", defined on line 446. # Proposals **Proposal 5 was selected for implementation.** ## Proposal 0: No Change Detection The baseline, where computations are performed on everything regardless of whether it changed. **Type:** Conservative **Pros:** - already implemented - will never miss events - no overhead **Cons:** - tons of repeated work - doesn't allow users to avoid repeating work (or monitoring for other changes) ## Proposal 1: Earlier-This-Tick Change Detection The current approach as of Bevy 0.4. Flags are set, and then flushed at the end of each frame. **Type:** Flaky **Pros:** - already implemented - simple to understand - low memory overhead (2 bits per component) - low time overhead (clear every flag once per frame) **Cons:** - misses systems based on ordering - systems that don't run every frame miss changes - duplicates detection when looping - can lead to unresolvable circular dependencies ## Proposal 2: Two-Tick Change Detection Flags persist for two frames, using a double-buffer system identical to that used in events. A change is observed if it is found in either the current frame's list of changes or the previous frame's. **Type:** Conservative **Pros:** - easy to understand - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - can result in a great deal of duplicated work - systems that don't run every frame miss changes - duplicates detection when looping ## Proposal 3: Last-Tick Change Detection Flags persist for two frames, using a double-buffer system identical to that used in events. A change is observed if it is found in the previous frame's list of changes. **Type:** Exact **Pros:** - exact - easy to understand - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - change detection is always delayed, possibly causing painful chained delays - systems that don't run every frame miss changes - duplicates detection when looping ## Proposal 4: Flag-Doubling Change Detection Combine Proposal 2 and Proposal 3. Differentiate between `JustChanged` (current behavior) and `Changed` (Proposal 3). Pack this data into the flags according to [this implementation proposal](https://github.com/bevyengine/bevy/issues/68#issuecomment-769174804). **Type:** Flaky + Exact **Pros:** - allows users to acc - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - users must specify the type of change detection required - still quite fragile to system ordering effects when using the flaky `JustChanged` form - cannot get immediate + exact results - systems that don't run every frame miss changes - duplicates detection when looping ## [SELECTED] Proposal 5: Generation-Counter Change Detection A global counter is increased after each system is run. Each component saves the time of last mutation, and each system saves the time of last execution. Mutation is detected when the component's counter is greater than the system's counter. Discussed [here](https://github.com/bevyengine/bevy/issues/68#issuecomment-769174804). How to handle addition detection is unsolved; the current proposal is to use the highest bit of the counter as in proposal 1. **Type:** Exact (for mutations), flaky (for additions) **Pros:** - low time overhead (set component counter on access, set system counter after execution) - robust to systems that don't run every frame - robust to systems that loop **Cons:** - moderately complex implementation - must be modified as systems are inserted dynamically - medium memory overhead (4 bytes per component + system) - unsolved addition detection ## Proposal 6: System-Data Change Detection For each system, track which system's changes it has seen. This approach is only worth fully designing and implementing if Proposal 5 fails in some way. **Type:** Exact **Pros:** - exact - conceptually simple **Cons:** - requires storing data on each system - implementation is complex - must be modified as systems are inserted dynamically ## Proposal 7: Total-Order Change Detection Discussed [here](https://github.com/bevyengine/bevy/issues/68#issuecomment-754326523). This proposal is somewhat complicated by the new scheduler, but I believe it should still be conceptually feasible. This approach is only worth fully designing and implementing if Proposal 5 fails in some way. **Type:** Exact **Pros:** - exact - efficient data storage relative to other exact proposals **Cons:** - requires access to the scheduler - complex implementation and difficulty grokking - must be modified as systems are inserted dynamically # Tests - We will need to verify properties 1, 2, 3, 7 and 8. Priority: 1 > 2 = 3 > 8 > 7 - Ideally we can use identical user-facing syntax for all proposals, allowing us to re-use the same syntax for each. - When writing tests, we need to carefully specify order using explicit dependencies. - These tests will need to be duplicated for both components and resources. - We need to be sure to handle cases where ambiguous system orders exist. `changing_system` is always the system that makes the changes, and `detecting_system` always detects the changes. The component / resource changed will be simple boolean wrapper structs. ## Basic Added / Mutated / Changed 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs before `detecting_system` - verify at the end of tick 2 ## At Least Once 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs after `detecting_system` - verify at the end of tick 2 ## At Most Once 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs once before `detecting_system` - increment a counter based on the number of changes detected - verify at the end of tick 2 ## Fast Detection 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs before `detecting_system` - verify at the end of tick 1 ## Ambiguous System Ordering Robustness 2 x 3 x 2 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs [before/after] `detecting_system` in tick 1 - `changing_system` runs [after/before] `detecting_system` in tick 2 ## System Pausing 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs in tick 1, then is disabled by run criteria - `detecting_system` is disabled by run criteria until it is run once during tick 3 - verify at the end of tick 3 ## Addition Causes Mutation 2 design: - Resources vs. Components - `adding_system_1` adds a component / resource - `adding system_2` adds the same component / resource - verify the `Mutated` flag at the end of the tick - verify the `Added` flag at the end of the tick First check tests for: https://github.com/bevyengine/bevy/issues/333 Second check tests for: https://github.com/bevyengine/bevy/issues/1443 ## Changes Made By Commands - `adding_system` runs in Update in tick 1, and sends a command to add a component - `detecting_system` runs in Update in tick 1 and 2, after `adding_system` - We can't detect the changes in tick 1, since they haven't been processed yet - If we were to track these changes as being emitted by `adding_system`, we can't detect the changes in tick 2 either, since `detecting_system` has already run once after `adding_system` :( # Benchmarks See: [general advice](https://github.com/bevyengine/bevy/blob/master/docs/profiling.md), [Criterion crate](https://github.com/bheisler/criterion.rs) There are several critical parameters to vary: 1. entity count (1 to 10^9) 2. fraction of entities that are changed (0% to 100%) 3. cost to perform work on changed entities, i.e. workload (1 ns to 1s) 1 and 2 should be varied between benchmark runs. 3 can be added on computationally. We want to measure: - memory cost - run time We should collect these measurements across several frames (100?) to reduce bootup effects and accurately measure the mean, variance and drift. Entity-component change detection is much more important to benchmark than resource change detection, due to the orders of magnitude higher number of pieces of data. No change detection at all should be included in benchmarks as a second control for cases where missing changes is unacceptable. ## Graphs 1. y: performance, x: log_10(entity count), color: proposal, facet: performance metric. Set cost to perform work to 0. 2. y: run time, x: cost to perform work, color: proposal, facet: fraction changed. Set number of entities to 10^6 3. y: memory, x: frames, color: proposal # Conclusions 1. Is the theoretical categorization of the proposals correct according to our tests? 2. How does the performance of the proposals compare without any load? 3. How does the performance of the proposals compare with realistic loads? 4. At what workload does more exact change tracking become worth the (presumably) higher overhead? 5. When does adding change-detection to save on work become worthwhile? 6. Is there enough divergence in performance between the best solutions in each class to ship more than one change-tracking solution? # Implementation Plan 1. Write a test suite. 2. Verify that tests fail for existing approach. 3. Write a benchmark suite. 4. Get performance numbers for existing approach. 5. Implement, test and benchmark various solutions using a Git branch per proposal. 6. Create a draft PR with all solutions and present results to team. 7. Select a solution and replace existing change detection. Co-authored-by: Brice DAVIER <bricedavier@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2021-03-19 17:53:26 +00:00
assert_eq!(get_filtered::<Changed<A>>(&mut world), vec![e3, e1], "changed entities list should not change (although the order will due to archetype moves)");
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
Reliable change detection (#1471) # Problem Definition The current change tracking (via flags for both components and resources) fails to detect changes made by systems that are scheduled to run earlier in the frame than they are. This issue is discussed at length in [#68](https://github.com/bevyengine/bevy/issues/68) and [#54](https://github.com/bevyengine/bevy/issues/54). This is very much a draft PR, and contributions are welcome and needed. # Criteria 1. Each change is detected at least once, no matter the ordering. 2. Each change is detected at most once, no matter the ordering. 3. Changes should be detected the same frame that they are made. 4. Competitive ergonomics. Ideally does not require opting-in. 5. Low CPU overhead of computation. 6. Memory efficient. This must not increase over time, except where the number of entities / resources does. 7. Changes should not be lost for systems that don't run. 8. A frame needs to act as a pure function. Given the same set of entities / components it needs to produce the same end state without side-effects. **Exact** change-tracking proposals satisfy criteria 1 and 2. **Conservative** change-tracking proposals satisfy criteria 1 but not 2. **Flaky** change tracking proposals satisfy criteria 2 but not 1. # Code Base Navigation There are three types of flags: - `Added`: A piece of data was added to an entity / `Resources`. - `Mutated`: A piece of data was able to be modified, because its `DerefMut` was accessed - `Changed`: The bitwise OR of `Added` and `Changed` The special behavior of `ChangedRes`, with respect to the scheduler is being removed in [#1313](https://github.com/bevyengine/bevy/pull/1313) and does not need to be reproduced. `ChangedRes` and friends can be found in "bevy_ecs/core/resources/resource_query.rs". The `Flags` trait for Components can be found in "bevy_ecs/core/query.rs". `ComponentFlags` are stored in "bevy_ecs/core/archetypes.rs", defined on line 446. # Proposals **Proposal 5 was selected for implementation.** ## Proposal 0: No Change Detection The baseline, where computations are performed on everything regardless of whether it changed. **Type:** Conservative **Pros:** - already implemented - will never miss events - no overhead **Cons:** - tons of repeated work - doesn't allow users to avoid repeating work (or monitoring for other changes) ## Proposal 1: Earlier-This-Tick Change Detection The current approach as of Bevy 0.4. Flags are set, and then flushed at the end of each frame. **Type:** Flaky **Pros:** - already implemented - simple to understand - low memory overhead (2 bits per component) - low time overhead (clear every flag once per frame) **Cons:** - misses systems based on ordering - systems that don't run every frame miss changes - duplicates detection when looping - can lead to unresolvable circular dependencies ## Proposal 2: Two-Tick Change Detection Flags persist for two frames, using a double-buffer system identical to that used in events. A change is observed if it is found in either the current frame's list of changes or the previous frame's. **Type:** Conservative **Pros:** - easy to understand - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - can result in a great deal of duplicated work - systems that don't run every frame miss changes - duplicates detection when looping ## Proposal 3: Last-Tick Change Detection Flags persist for two frames, using a double-buffer system identical to that used in events. A change is observed if it is found in the previous frame's list of changes. **Type:** Exact **Pros:** - exact - easy to understand - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - change detection is always delayed, possibly causing painful chained delays - systems that don't run every frame miss changes - duplicates detection when looping ## Proposal 4: Flag-Doubling Change Detection Combine Proposal 2 and Proposal 3. Differentiate between `JustChanged` (current behavior) and `Changed` (Proposal 3). Pack this data into the flags according to [this implementation proposal](https://github.com/bevyengine/bevy/issues/68#issuecomment-769174804). **Type:** Flaky + Exact **Pros:** - allows users to acc - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - users must specify the type of change detection required - still quite fragile to system ordering effects when using the flaky `JustChanged` form - cannot get immediate + exact results - systems that don't run every frame miss changes - duplicates detection when looping ## [SELECTED] Proposal 5: Generation-Counter Change Detection A global counter is increased after each system is run. Each component saves the time of last mutation, and each system saves the time of last execution. Mutation is detected when the component's counter is greater than the system's counter. Discussed [here](https://github.com/bevyengine/bevy/issues/68#issuecomment-769174804). How to handle addition detection is unsolved; the current proposal is to use the highest bit of the counter as in proposal 1. **Type:** Exact (for mutations), flaky (for additions) **Pros:** - low time overhead (set component counter on access, set system counter after execution) - robust to systems that don't run every frame - robust to systems that loop **Cons:** - moderately complex implementation - must be modified as systems are inserted dynamically - medium memory overhead (4 bytes per component + system) - unsolved addition detection ## Proposal 6: System-Data Change Detection For each system, track which system's changes it has seen. This approach is only worth fully designing and implementing if Proposal 5 fails in some way. **Type:** Exact **Pros:** - exact - conceptually simple **Cons:** - requires storing data on each system - implementation is complex - must be modified as systems are inserted dynamically ## Proposal 7: Total-Order Change Detection Discussed [here](https://github.com/bevyengine/bevy/issues/68#issuecomment-754326523). This proposal is somewhat complicated by the new scheduler, but I believe it should still be conceptually feasible. This approach is only worth fully designing and implementing if Proposal 5 fails in some way. **Type:** Exact **Pros:** - exact - efficient data storage relative to other exact proposals **Cons:** - requires access to the scheduler - complex implementation and difficulty grokking - must be modified as systems are inserted dynamically # Tests - We will need to verify properties 1, 2, 3, 7 and 8. Priority: 1 > 2 = 3 > 8 > 7 - Ideally we can use identical user-facing syntax for all proposals, allowing us to re-use the same syntax for each. - When writing tests, we need to carefully specify order using explicit dependencies. - These tests will need to be duplicated for both components and resources. - We need to be sure to handle cases where ambiguous system orders exist. `changing_system` is always the system that makes the changes, and `detecting_system` always detects the changes. The component / resource changed will be simple boolean wrapper structs. ## Basic Added / Mutated / Changed 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs before `detecting_system` - verify at the end of tick 2 ## At Least Once 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs after `detecting_system` - verify at the end of tick 2 ## At Most Once 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs once before `detecting_system` - increment a counter based on the number of changes detected - verify at the end of tick 2 ## Fast Detection 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs before `detecting_system` - verify at the end of tick 1 ## Ambiguous System Ordering Robustness 2 x 3 x 2 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs [before/after] `detecting_system` in tick 1 - `changing_system` runs [after/before] `detecting_system` in tick 2 ## System Pausing 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs in tick 1, then is disabled by run criteria - `detecting_system` is disabled by run criteria until it is run once during tick 3 - verify at the end of tick 3 ## Addition Causes Mutation 2 design: - Resources vs. Components - `adding_system_1` adds a component / resource - `adding system_2` adds the same component / resource - verify the `Mutated` flag at the end of the tick - verify the `Added` flag at the end of the tick First check tests for: https://github.com/bevyengine/bevy/issues/333 Second check tests for: https://github.com/bevyengine/bevy/issues/1443 ## Changes Made By Commands - `adding_system` runs in Update in tick 1, and sends a command to add a component - `detecting_system` runs in Update in tick 1 and 2, after `adding_system` - We can't detect the changes in tick 1, since they haven't been processed yet - If we were to track these changes as being emitted by `adding_system`, we can't detect the changes in tick 2 either, since `detecting_system` has already run once after `adding_system` :( # Benchmarks See: [general advice](https://github.com/bevyengine/bevy/blob/master/docs/profiling.md), [Criterion crate](https://github.com/bheisler/criterion.rs) There are several critical parameters to vary: 1. entity count (1 to 10^9) 2. fraction of entities that are changed (0% to 100%) 3. cost to perform work on changed entities, i.e. workload (1 ns to 1s) 1 and 2 should be varied between benchmark runs. 3 can be added on computationally. We want to measure: - memory cost - run time We should collect these measurements across several frames (100?) to reduce bootup effects and accurately measure the mean, variance and drift. Entity-component change detection is much more important to benchmark than resource change detection, due to the orders of magnitude higher number of pieces of data. No change detection at all should be included in benchmarks as a second control for cases where missing changes is unacceptable. ## Graphs 1. y: performance, x: log_10(entity count), color: proposal, facet: performance metric. Set cost to perform work to 0. 2. y: run time, x: cost to perform work, color: proposal, facet: fraction changed. Set number of entities to 10^6 3. y: memory, x: frames, color: proposal # Conclusions 1. Is the theoretical categorization of the proposals correct according to our tests? 2. How does the performance of the proposals compare without any load? 3. How does the performance of the proposals compare with realistic loads? 4. At what workload does more exact change tracking become worth the (presumably) higher overhead? 5. When does adding change-detection to save on work become worthwhile? 6. Is there enough divergence in performance between the best solutions in each class to ship more than one change-tracking solution? # Implementation Plan 1. Write a test suite. 2. Verify that tests fail for existing approach. 3. Write a benchmark suite. 4. Get performance numbers for existing approach. 5. Implement, test and benchmark various solutions using a Git branch per proposal. 6. Create a draft PR with all solutions and present results to team. 7. Select a solution and replace existing change detection. Co-authored-by: Brice DAVIER <bricedavier@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2021-03-19 17:53:26 +00:00
// spawning a new A entity should not change existing changed state
Accept Bundles for insert and remove. Deprecate insert/remove_bundle (#6039) # Objective Take advantage of the "impl Bundle for Component" changes in #2975 / add the follow up changes discussed there. ## Solution - Change `insert` and `remove` to accept a Bundle instead of a Component (for both Commands and World) - Deprecate `insert_bundle`, `remove_bundle`, and `remove_bundle_intersection` - Add `remove_intersection` --- ## Changelog - Change `insert` and `remove` now accept a Bundle instead of a Component (for both Commands and World) - `insert_bundle` and `remove_bundle` are deprecated ## Migration Guide Replace `insert_bundle` with `insert`: ```rust // Old (0.8) commands.spawn().insert_bundle(SomeBundle::default()); // New (0.9) commands.spawn().insert(SomeBundle::default()); ``` Replace `remove_bundle` with `remove`: ```rust // Old (0.8) commands.entity(some_entity).remove_bundle::<SomeBundle>(); // New (0.9) commands.entity(some_entity).remove::<SomeBundle>(); ``` Replace `remove_bundle_intersection` with `remove_intersection`: ```rust // Old (0.8) world.entity_mut(some_entity).remove_bundle_intersection::<SomeBundle>(); // New (0.9) world.entity_mut(some_entity).remove_intersection::<SomeBundle>(); ``` Consider consolidating as many operations as possible to improve ergonomics and cut down on archetype moves: ```rust // Old (0.8) commands.spawn() .insert_bundle(SomeBundle::default()) .insert(SomeComponent); // New (0.9) - Option 1 commands.spawn().insert(( SomeBundle::default(), SomeComponent, )) // New (0.9) - Option 2 commands.spawn_bundle(( SomeBundle::default(), SomeComponent, )) ``` ## Next Steps Consider changing `spawn` to accept a bundle and deprecate `spawn_bundle`.
2022-09-21 21:47:53 +00:00
world.entity_mut(e1).insert((A(0), B(0)));
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
assert_eq!(
Reliable change detection (#1471) # Problem Definition The current change tracking (via flags for both components and resources) fails to detect changes made by systems that are scheduled to run earlier in the frame than they are. This issue is discussed at length in [#68](https://github.com/bevyengine/bevy/issues/68) and [#54](https://github.com/bevyengine/bevy/issues/54). This is very much a draft PR, and contributions are welcome and needed. # Criteria 1. Each change is detected at least once, no matter the ordering. 2. Each change is detected at most once, no matter the ordering. 3. Changes should be detected the same frame that they are made. 4. Competitive ergonomics. Ideally does not require opting-in. 5. Low CPU overhead of computation. 6. Memory efficient. This must not increase over time, except where the number of entities / resources does. 7. Changes should not be lost for systems that don't run. 8. A frame needs to act as a pure function. Given the same set of entities / components it needs to produce the same end state without side-effects. **Exact** change-tracking proposals satisfy criteria 1 and 2. **Conservative** change-tracking proposals satisfy criteria 1 but not 2. **Flaky** change tracking proposals satisfy criteria 2 but not 1. # Code Base Navigation There are three types of flags: - `Added`: A piece of data was added to an entity / `Resources`. - `Mutated`: A piece of data was able to be modified, because its `DerefMut` was accessed - `Changed`: The bitwise OR of `Added` and `Changed` The special behavior of `ChangedRes`, with respect to the scheduler is being removed in [#1313](https://github.com/bevyengine/bevy/pull/1313) and does not need to be reproduced. `ChangedRes` and friends can be found in "bevy_ecs/core/resources/resource_query.rs". The `Flags` trait for Components can be found in "bevy_ecs/core/query.rs". `ComponentFlags` are stored in "bevy_ecs/core/archetypes.rs", defined on line 446. # Proposals **Proposal 5 was selected for implementation.** ## Proposal 0: No Change Detection The baseline, where computations are performed on everything regardless of whether it changed. **Type:** Conservative **Pros:** - already implemented - will never miss events - no overhead **Cons:** - tons of repeated work - doesn't allow users to avoid repeating work (or monitoring for other changes) ## Proposal 1: Earlier-This-Tick Change Detection The current approach as of Bevy 0.4. Flags are set, and then flushed at the end of each frame. **Type:** Flaky **Pros:** - already implemented - simple to understand - low memory overhead (2 bits per component) - low time overhead (clear every flag once per frame) **Cons:** - misses systems based on ordering - systems that don't run every frame miss changes - duplicates detection when looping - can lead to unresolvable circular dependencies ## Proposal 2: Two-Tick Change Detection Flags persist for two frames, using a double-buffer system identical to that used in events. A change is observed if it is found in either the current frame's list of changes or the previous frame's. **Type:** Conservative **Pros:** - easy to understand - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - can result in a great deal of duplicated work - systems that don't run every frame miss changes - duplicates detection when looping ## Proposal 3: Last-Tick Change Detection Flags persist for two frames, using a double-buffer system identical to that used in events. A change is observed if it is found in the previous frame's list of changes. **Type:** Exact **Pros:** - exact - easy to understand - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - change detection is always delayed, possibly causing painful chained delays - systems that don't run every frame miss changes - duplicates detection when looping ## Proposal 4: Flag-Doubling Change Detection Combine Proposal 2 and Proposal 3. Differentiate between `JustChanged` (current behavior) and `Changed` (Proposal 3). Pack this data into the flags according to [this implementation proposal](https://github.com/bevyengine/bevy/issues/68#issuecomment-769174804). **Type:** Flaky + Exact **Pros:** - allows users to acc - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - users must specify the type of change detection required - still quite fragile to system ordering effects when using the flaky `JustChanged` form - cannot get immediate + exact results - systems that don't run every frame miss changes - duplicates detection when looping ## [SELECTED] Proposal 5: Generation-Counter Change Detection A global counter is increased after each system is run. Each component saves the time of last mutation, and each system saves the time of last execution. Mutation is detected when the component's counter is greater than the system's counter. Discussed [here](https://github.com/bevyengine/bevy/issues/68#issuecomment-769174804). How to handle addition detection is unsolved; the current proposal is to use the highest bit of the counter as in proposal 1. **Type:** Exact (for mutations), flaky (for additions) **Pros:** - low time overhead (set component counter on access, set system counter after execution) - robust to systems that don't run every frame - robust to systems that loop **Cons:** - moderately complex implementation - must be modified as systems are inserted dynamically - medium memory overhead (4 bytes per component + system) - unsolved addition detection ## Proposal 6: System-Data Change Detection For each system, track which system's changes it has seen. This approach is only worth fully designing and implementing if Proposal 5 fails in some way. **Type:** Exact **Pros:** - exact - conceptually simple **Cons:** - requires storing data on each system - implementation is complex - must be modified as systems are inserted dynamically ## Proposal 7: Total-Order Change Detection Discussed [here](https://github.com/bevyengine/bevy/issues/68#issuecomment-754326523). This proposal is somewhat complicated by the new scheduler, but I believe it should still be conceptually feasible. This approach is only worth fully designing and implementing if Proposal 5 fails in some way. **Type:** Exact **Pros:** - exact - efficient data storage relative to other exact proposals **Cons:** - requires access to the scheduler - complex implementation and difficulty grokking - must be modified as systems are inserted dynamically # Tests - We will need to verify properties 1, 2, 3, 7 and 8. Priority: 1 > 2 = 3 > 8 > 7 - Ideally we can use identical user-facing syntax for all proposals, allowing us to re-use the same syntax for each. - When writing tests, we need to carefully specify order using explicit dependencies. - These tests will need to be duplicated for both components and resources. - We need to be sure to handle cases where ambiguous system orders exist. `changing_system` is always the system that makes the changes, and `detecting_system` always detects the changes. The component / resource changed will be simple boolean wrapper structs. ## Basic Added / Mutated / Changed 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs before `detecting_system` - verify at the end of tick 2 ## At Least Once 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs after `detecting_system` - verify at the end of tick 2 ## At Most Once 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs once before `detecting_system` - increment a counter based on the number of changes detected - verify at the end of tick 2 ## Fast Detection 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs before `detecting_system` - verify at the end of tick 1 ## Ambiguous System Ordering Robustness 2 x 3 x 2 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs [before/after] `detecting_system` in tick 1 - `changing_system` runs [after/before] `detecting_system` in tick 2 ## System Pausing 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs in tick 1, then is disabled by run criteria - `detecting_system` is disabled by run criteria until it is run once during tick 3 - verify at the end of tick 3 ## Addition Causes Mutation 2 design: - Resources vs. Components - `adding_system_1` adds a component / resource - `adding system_2` adds the same component / resource - verify the `Mutated` flag at the end of the tick - verify the `Added` flag at the end of the tick First check tests for: https://github.com/bevyengine/bevy/issues/333 Second check tests for: https://github.com/bevyengine/bevy/issues/1443 ## Changes Made By Commands - `adding_system` runs in Update in tick 1, and sends a command to add a component - `detecting_system` runs in Update in tick 1 and 2, after `adding_system` - We can't detect the changes in tick 1, since they haven't been processed yet - If we were to track these changes as being emitted by `adding_system`, we can't detect the changes in tick 2 either, since `detecting_system` has already run once after `adding_system` :( # Benchmarks See: [general advice](https://github.com/bevyengine/bevy/blob/master/docs/profiling.md), [Criterion crate](https://github.com/bheisler/criterion.rs) There are several critical parameters to vary: 1. entity count (1 to 10^9) 2. fraction of entities that are changed (0% to 100%) 3. cost to perform work on changed entities, i.e. workload (1 ns to 1s) 1 and 2 should be varied between benchmark runs. 3 can be added on computationally. We want to measure: - memory cost - run time We should collect these measurements across several frames (100?) to reduce bootup effects and accurately measure the mean, variance and drift. Entity-component change detection is much more important to benchmark than resource change detection, due to the orders of magnitude higher number of pieces of data. No change detection at all should be included in benchmarks as a second control for cases where missing changes is unacceptable. ## Graphs 1. y: performance, x: log_10(entity count), color: proposal, facet: performance metric. Set cost to perform work to 0. 2. y: run time, x: cost to perform work, color: proposal, facet: fraction changed. Set number of entities to 10^6 3. y: memory, x: frames, color: proposal # Conclusions 1. Is the theoretical categorization of the proposals correct according to our tests? 2. How does the performance of the proposals compare without any load? 3. How does the performance of the proposals compare with realistic loads? 4. At what workload does more exact change tracking become worth the (presumably) higher overhead? 5. When does adding change-detection to save on work become worthwhile? 6. Is there enough divergence in performance between the best solutions in each class to ship more than one change-tracking solution? # Implementation Plan 1. Write a test suite. 2. Verify that tests fail for existing approach. 3. Write a benchmark suite. 4. Get performance numbers for existing approach. 5. Implement, test and benchmark various solutions using a Git branch per proposal. 6. Create a draft PR with all solutions and present results to team. 7. Select a solution and replace existing change detection. Co-authored-by: Brice DAVIER <bricedavier@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2021-03-19 17:53:26 +00:00
get_filtered::<Changed<A>>(&mut world),
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
vec![e3, e1],
"changed entities list should not change"
);
Reliable change detection (#1471) # Problem Definition The current change tracking (via flags for both components and resources) fails to detect changes made by systems that are scheduled to run earlier in the frame than they are. This issue is discussed at length in [#68](https://github.com/bevyengine/bevy/issues/68) and [#54](https://github.com/bevyengine/bevy/issues/54). This is very much a draft PR, and contributions are welcome and needed. # Criteria 1. Each change is detected at least once, no matter the ordering. 2. Each change is detected at most once, no matter the ordering. 3. Changes should be detected the same frame that they are made. 4. Competitive ergonomics. Ideally does not require opting-in. 5. Low CPU overhead of computation. 6. Memory efficient. This must not increase over time, except where the number of entities / resources does. 7. Changes should not be lost for systems that don't run. 8. A frame needs to act as a pure function. Given the same set of entities / components it needs to produce the same end state without side-effects. **Exact** change-tracking proposals satisfy criteria 1 and 2. **Conservative** change-tracking proposals satisfy criteria 1 but not 2. **Flaky** change tracking proposals satisfy criteria 2 but not 1. # Code Base Navigation There are three types of flags: - `Added`: A piece of data was added to an entity / `Resources`. - `Mutated`: A piece of data was able to be modified, because its `DerefMut` was accessed - `Changed`: The bitwise OR of `Added` and `Changed` The special behavior of `ChangedRes`, with respect to the scheduler is being removed in [#1313](https://github.com/bevyengine/bevy/pull/1313) and does not need to be reproduced. `ChangedRes` and friends can be found in "bevy_ecs/core/resources/resource_query.rs". The `Flags` trait for Components can be found in "bevy_ecs/core/query.rs". `ComponentFlags` are stored in "bevy_ecs/core/archetypes.rs", defined on line 446. # Proposals **Proposal 5 was selected for implementation.** ## Proposal 0: No Change Detection The baseline, where computations are performed on everything regardless of whether it changed. **Type:** Conservative **Pros:** - already implemented - will never miss events - no overhead **Cons:** - tons of repeated work - doesn't allow users to avoid repeating work (or monitoring for other changes) ## Proposal 1: Earlier-This-Tick Change Detection The current approach as of Bevy 0.4. Flags are set, and then flushed at the end of each frame. **Type:** Flaky **Pros:** - already implemented - simple to understand - low memory overhead (2 bits per component) - low time overhead (clear every flag once per frame) **Cons:** - misses systems based on ordering - systems that don't run every frame miss changes - duplicates detection when looping - can lead to unresolvable circular dependencies ## Proposal 2: Two-Tick Change Detection Flags persist for two frames, using a double-buffer system identical to that used in events. A change is observed if it is found in either the current frame's list of changes or the previous frame's. **Type:** Conservative **Pros:** - easy to understand - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - can result in a great deal of duplicated work - systems that don't run every frame miss changes - duplicates detection when looping ## Proposal 3: Last-Tick Change Detection Flags persist for two frames, using a double-buffer system identical to that used in events. A change is observed if it is found in the previous frame's list of changes. **Type:** Exact **Pros:** - exact - easy to understand - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - change detection is always delayed, possibly causing painful chained delays - systems that don't run every frame miss changes - duplicates detection when looping ## Proposal 4: Flag-Doubling Change Detection Combine Proposal 2 and Proposal 3. Differentiate between `JustChanged` (current behavior) and `Changed` (Proposal 3). Pack this data into the flags according to [this implementation proposal](https://github.com/bevyengine/bevy/issues/68#issuecomment-769174804). **Type:** Flaky + Exact **Pros:** - allows users to acc - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - users must specify the type of change detection required - still quite fragile to system ordering effects when using the flaky `JustChanged` form - cannot get immediate + exact results - systems that don't run every frame miss changes - duplicates detection when looping ## [SELECTED] Proposal 5: Generation-Counter Change Detection A global counter is increased after each system is run. Each component saves the time of last mutation, and each system saves the time of last execution. Mutation is detected when the component's counter is greater than the system's counter. Discussed [here](https://github.com/bevyengine/bevy/issues/68#issuecomment-769174804). How to handle addition detection is unsolved; the current proposal is to use the highest bit of the counter as in proposal 1. **Type:** Exact (for mutations), flaky (for additions) **Pros:** - low time overhead (set component counter on access, set system counter after execution) - robust to systems that don't run every frame - robust to systems that loop **Cons:** - moderately complex implementation - must be modified as systems are inserted dynamically - medium memory overhead (4 bytes per component + system) - unsolved addition detection ## Proposal 6: System-Data Change Detection For each system, track which system's changes it has seen. This approach is only worth fully designing and implementing if Proposal 5 fails in some way. **Type:** Exact **Pros:** - exact - conceptually simple **Cons:** - requires storing data on each system - implementation is complex - must be modified as systems are inserted dynamically ## Proposal 7: Total-Order Change Detection Discussed [here](https://github.com/bevyengine/bevy/issues/68#issuecomment-754326523). This proposal is somewhat complicated by the new scheduler, but I believe it should still be conceptually feasible. This approach is only worth fully designing and implementing if Proposal 5 fails in some way. **Type:** Exact **Pros:** - exact - efficient data storage relative to other exact proposals **Cons:** - requires access to the scheduler - complex implementation and difficulty grokking - must be modified as systems are inserted dynamically # Tests - We will need to verify properties 1, 2, 3, 7 and 8. Priority: 1 > 2 = 3 > 8 > 7 - Ideally we can use identical user-facing syntax for all proposals, allowing us to re-use the same syntax for each. - When writing tests, we need to carefully specify order using explicit dependencies. - These tests will need to be duplicated for both components and resources. - We need to be sure to handle cases where ambiguous system orders exist. `changing_system` is always the system that makes the changes, and `detecting_system` always detects the changes. The component / resource changed will be simple boolean wrapper structs. ## Basic Added / Mutated / Changed 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs before `detecting_system` - verify at the end of tick 2 ## At Least Once 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs after `detecting_system` - verify at the end of tick 2 ## At Most Once 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs once before `detecting_system` - increment a counter based on the number of changes detected - verify at the end of tick 2 ## Fast Detection 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs before `detecting_system` - verify at the end of tick 1 ## Ambiguous System Ordering Robustness 2 x 3 x 2 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs [before/after] `detecting_system` in tick 1 - `changing_system` runs [after/before] `detecting_system` in tick 2 ## System Pausing 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs in tick 1, then is disabled by run criteria - `detecting_system` is disabled by run criteria until it is run once during tick 3 - verify at the end of tick 3 ## Addition Causes Mutation 2 design: - Resources vs. Components - `adding_system_1` adds a component / resource - `adding system_2` adds the same component / resource - verify the `Mutated` flag at the end of the tick - verify the `Added` flag at the end of the tick First check tests for: https://github.com/bevyengine/bevy/issues/333 Second check tests for: https://github.com/bevyengine/bevy/issues/1443 ## Changes Made By Commands - `adding_system` runs in Update in tick 1, and sends a command to add a component - `detecting_system` runs in Update in tick 1 and 2, after `adding_system` - We can't detect the changes in tick 1, since they haven't been processed yet - If we were to track these changes as being emitted by `adding_system`, we can't detect the changes in tick 2 either, since `detecting_system` has already run once after `adding_system` :( # Benchmarks See: [general advice](https://github.com/bevyengine/bevy/blob/master/docs/profiling.md), [Criterion crate](https://github.com/bheisler/criterion.rs) There are several critical parameters to vary: 1. entity count (1 to 10^9) 2. fraction of entities that are changed (0% to 100%) 3. cost to perform work on changed entities, i.e. workload (1 ns to 1s) 1 and 2 should be varied between benchmark runs. 3 can be added on computationally. We want to measure: - memory cost - run time We should collect these measurements across several frames (100?) to reduce bootup effects and accurately measure the mean, variance and drift. Entity-component change detection is much more important to benchmark than resource change detection, due to the orders of magnitude higher number of pieces of data. No change detection at all should be included in benchmarks as a second control for cases where missing changes is unacceptable. ## Graphs 1. y: performance, x: log_10(entity count), color: proposal, facet: performance metric. Set cost to perform work to 0. 2. y: run time, x: cost to perform work, color: proposal, facet: fraction changed. Set number of entities to 10^6 3. y: memory, x: frames, color: proposal # Conclusions 1. Is the theoretical categorization of the proposals correct according to our tests? 2. How does the performance of the proposals compare without any load? 3. How does the performance of the proposals compare with realistic loads? 4. At what workload does more exact change tracking become worth the (presumably) higher overhead? 5. When does adding change-detection to save on work become worthwhile? 6. Is there enough divergence in performance between the best solutions in each class to ship more than one change-tracking solution? # Implementation Plan 1. Write a test suite. 2. Verify that tests fail for existing approach. 3. Write a benchmark suite. 4. Get performance numbers for existing approach. 5. Implement, test and benchmark various solutions using a Git branch per proposal. 6. Create a draft PR with all solutions and present results to team. 7. Select a solution and replace existing change detection. Co-authored-by: Brice DAVIER <bricedavier@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2021-03-19 17:53:26 +00:00
// removing an unchanged entity should not change changed state
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
assert!(world.despawn(e2));
assert_eq!(
Reliable change detection (#1471) # Problem Definition The current change tracking (via flags for both components and resources) fails to detect changes made by systems that are scheduled to run earlier in the frame than they are. This issue is discussed at length in [#68](https://github.com/bevyengine/bevy/issues/68) and [#54](https://github.com/bevyengine/bevy/issues/54). This is very much a draft PR, and contributions are welcome and needed. # Criteria 1. Each change is detected at least once, no matter the ordering. 2. Each change is detected at most once, no matter the ordering. 3. Changes should be detected the same frame that they are made. 4. Competitive ergonomics. Ideally does not require opting-in. 5. Low CPU overhead of computation. 6. Memory efficient. This must not increase over time, except where the number of entities / resources does. 7. Changes should not be lost for systems that don't run. 8. A frame needs to act as a pure function. Given the same set of entities / components it needs to produce the same end state without side-effects. **Exact** change-tracking proposals satisfy criteria 1 and 2. **Conservative** change-tracking proposals satisfy criteria 1 but not 2. **Flaky** change tracking proposals satisfy criteria 2 but not 1. # Code Base Navigation There are three types of flags: - `Added`: A piece of data was added to an entity / `Resources`. - `Mutated`: A piece of data was able to be modified, because its `DerefMut` was accessed - `Changed`: The bitwise OR of `Added` and `Changed` The special behavior of `ChangedRes`, with respect to the scheduler is being removed in [#1313](https://github.com/bevyengine/bevy/pull/1313) and does not need to be reproduced. `ChangedRes` and friends can be found in "bevy_ecs/core/resources/resource_query.rs". The `Flags` trait for Components can be found in "bevy_ecs/core/query.rs". `ComponentFlags` are stored in "bevy_ecs/core/archetypes.rs", defined on line 446. # Proposals **Proposal 5 was selected for implementation.** ## Proposal 0: No Change Detection The baseline, where computations are performed on everything regardless of whether it changed. **Type:** Conservative **Pros:** - already implemented - will never miss events - no overhead **Cons:** - tons of repeated work - doesn't allow users to avoid repeating work (or monitoring for other changes) ## Proposal 1: Earlier-This-Tick Change Detection The current approach as of Bevy 0.4. Flags are set, and then flushed at the end of each frame. **Type:** Flaky **Pros:** - already implemented - simple to understand - low memory overhead (2 bits per component) - low time overhead (clear every flag once per frame) **Cons:** - misses systems based on ordering - systems that don't run every frame miss changes - duplicates detection when looping - can lead to unresolvable circular dependencies ## Proposal 2: Two-Tick Change Detection Flags persist for two frames, using a double-buffer system identical to that used in events. A change is observed if it is found in either the current frame's list of changes or the previous frame's. **Type:** Conservative **Pros:** - easy to understand - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - can result in a great deal of duplicated work - systems that don't run every frame miss changes - duplicates detection when looping ## Proposal 3: Last-Tick Change Detection Flags persist for two frames, using a double-buffer system identical to that used in events. A change is observed if it is found in the previous frame's list of changes. **Type:** Exact **Pros:** - exact - easy to understand - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - change detection is always delayed, possibly causing painful chained delays - systems that don't run every frame miss changes - duplicates detection when looping ## Proposal 4: Flag-Doubling Change Detection Combine Proposal 2 and Proposal 3. Differentiate between `JustChanged` (current behavior) and `Changed` (Proposal 3). Pack this data into the flags according to [this implementation proposal](https://github.com/bevyengine/bevy/issues/68#issuecomment-769174804). **Type:** Flaky + Exact **Pros:** - allows users to acc - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - users must specify the type of change detection required - still quite fragile to system ordering effects when using the flaky `JustChanged` form - cannot get immediate + exact results - systems that don't run every frame miss changes - duplicates detection when looping ## [SELECTED] Proposal 5: Generation-Counter Change Detection A global counter is increased after each system is run. Each component saves the time of last mutation, and each system saves the time of last execution. Mutation is detected when the component's counter is greater than the system's counter. Discussed [here](https://github.com/bevyengine/bevy/issues/68#issuecomment-769174804). How to handle addition detection is unsolved; the current proposal is to use the highest bit of the counter as in proposal 1. **Type:** Exact (for mutations), flaky (for additions) **Pros:** - low time overhead (set component counter on access, set system counter after execution) - robust to systems that don't run every frame - robust to systems that loop **Cons:** - moderately complex implementation - must be modified as systems are inserted dynamically - medium memory overhead (4 bytes per component + system) - unsolved addition detection ## Proposal 6: System-Data Change Detection For each system, track which system's changes it has seen. This approach is only worth fully designing and implementing if Proposal 5 fails in some way. **Type:** Exact **Pros:** - exact - conceptually simple **Cons:** - requires storing data on each system - implementation is complex - must be modified as systems are inserted dynamically ## Proposal 7: Total-Order Change Detection Discussed [here](https://github.com/bevyengine/bevy/issues/68#issuecomment-754326523). This proposal is somewhat complicated by the new scheduler, but I believe it should still be conceptually feasible. This approach is only worth fully designing and implementing if Proposal 5 fails in some way. **Type:** Exact **Pros:** - exact - efficient data storage relative to other exact proposals **Cons:** - requires access to the scheduler - complex implementation and difficulty grokking - must be modified as systems are inserted dynamically # Tests - We will need to verify properties 1, 2, 3, 7 and 8. Priority: 1 > 2 = 3 > 8 > 7 - Ideally we can use identical user-facing syntax for all proposals, allowing us to re-use the same syntax for each. - When writing tests, we need to carefully specify order using explicit dependencies. - These tests will need to be duplicated for both components and resources. - We need to be sure to handle cases where ambiguous system orders exist. `changing_system` is always the system that makes the changes, and `detecting_system` always detects the changes. The component / resource changed will be simple boolean wrapper structs. ## Basic Added / Mutated / Changed 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs before `detecting_system` - verify at the end of tick 2 ## At Least Once 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs after `detecting_system` - verify at the end of tick 2 ## At Most Once 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs once before `detecting_system` - increment a counter based on the number of changes detected - verify at the end of tick 2 ## Fast Detection 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs before `detecting_system` - verify at the end of tick 1 ## Ambiguous System Ordering Robustness 2 x 3 x 2 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs [before/after] `detecting_system` in tick 1 - `changing_system` runs [after/before] `detecting_system` in tick 2 ## System Pausing 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs in tick 1, then is disabled by run criteria - `detecting_system` is disabled by run criteria until it is run once during tick 3 - verify at the end of tick 3 ## Addition Causes Mutation 2 design: - Resources vs. Components - `adding_system_1` adds a component / resource - `adding system_2` adds the same component / resource - verify the `Mutated` flag at the end of the tick - verify the `Added` flag at the end of the tick First check tests for: https://github.com/bevyengine/bevy/issues/333 Second check tests for: https://github.com/bevyengine/bevy/issues/1443 ## Changes Made By Commands - `adding_system` runs in Update in tick 1, and sends a command to add a component - `detecting_system` runs in Update in tick 1 and 2, after `adding_system` - We can't detect the changes in tick 1, since they haven't been processed yet - If we were to track these changes as being emitted by `adding_system`, we can't detect the changes in tick 2 either, since `detecting_system` has already run once after `adding_system` :( # Benchmarks See: [general advice](https://github.com/bevyengine/bevy/blob/master/docs/profiling.md), [Criterion crate](https://github.com/bheisler/criterion.rs) There are several critical parameters to vary: 1. entity count (1 to 10^9) 2. fraction of entities that are changed (0% to 100%) 3. cost to perform work on changed entities, i.e. workload (1 ns to 1s) 1 and 2 should be varied between benchmark runs. 3 can be added on computationally. We want to measure: - memory cost - run time We should collect these measurements across several frames (100?) to reduce bootup effects and accurately measure the mean, variance and drift. Entity-component change detection is much more important to benchmark than resource change detection, due to the orders of magnitude higher number of pieces of data. No change detection at all should be included in benchmarks as a second control for cases where missing changes is unacceptable. ## Graphs 1. y: performance, x: log_10(entity count), color: proposal, facet: performance metric. Set cost to perform work to 0. 2. y: run time, x: cost to perform work, color: proposal, facet: fraction changed. Set number of entities to 10^6 3. y: memory, x: frames, color: proposal # Conclusions 1. Is the theoretical categorization of the proposals correct according to our tests? 2. How does the performance of the proposals compare without any load? 3. How does the performance of the proposals compare with realistic loads? 4. At what workload does more exact change tracking become worth the (presumably) higher overhead? 5. When does adding change-detection to save on work become worthwhile? 6. Is there enough divergence in performance between the best solutions in each class to ship more than one change-tracking solution? # Implementation Plan 1. Write a test suite. 2. Verify that tests fail for existing approach. 3. Write a benchmark suite. 4. Get performance numbers for existing approach. 5. Implement, test and benchmark various solutions using a Git branch per proposal. 6. Create a draft PR with all solutions and present results to team. 7. Select a solution and replace existing change detection. Co-authored-by: Brice DAVIER <bricedavier@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2021-03-19 17:53:26 +00:00
get_filtered::<Changed<A>>(&mut world),
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
vec![e3, e1],
"changed entities list should not change"
);
// removing a changed entity should remove it from enumeration
assert!(world.despawn(e1));
assert_eq!(
Reliable change detection (#1471) # Problem Definition The current change tracking (via flags for both components and resources) fails to detect changes made by systems that are scheduled to run earlier in the frame than they are. This issue is discussed at length in [#68](https://github.com/bevyengine/bevy/issues/68) and [#54](https://github.com/bevyengine/bevy/issues/54). This is very much a draft PR, and contributions are welcome and needed. # Criteria 1. Each change is detected at least once, no matter the ordering. 2. Each change is detected at most once, no matter the ordering. 3. Changes should be detected the same frame that they are made. 4. Competitive ergonomics. Ideally does not require opting-in. 5. Low CPU overhead of computation. 6. Memory efficient. This must not increase over time, except where the number of entities / resources does. 7. Changes should not be lost for systems that don't run. 8. A frame needs to act as a pure function. Given the same set of entities / components it needs to produce the same end state without side-effects. **Exact** change-tracking proposals satisfy criteria 1 and 2. **Conservative** change-tracking proposals satisfy criteria 1 but not 2. **Flaky** change tracking proposals satisfy criteria 2 but not 1. # Code Base Navigation There are three types of flags: - `Added`: A piece of data was added to an entity / `Resources`. - `Mutated`: A piece of data was able to be modified, because its `DerefMut` was accessed - `Changed`: The bitwise OR of `Added` and `Changed` The special behavior of `ChangedRes`, with respect to the scheduler is being removed in [#1313](https://github.com/bevyengine/bevy/pull/1313) and does not need to be reproduced. `ChangedRes` and friends can be found in "bevy_ecs/core/resources/resource_query.rs". The `Flags` trait for Components can be found in "bevy_ecs/core/query.rs". `ComponentFlags` are stored in "bevy_ecs/core/archetypes.rs", defined on line 446. # Proposals **Proposal 5 was selected for implementation.** ## Proposal 0: No Change Detection The baseline, where computations are performed on everything regardless of whether it changed. **Type:** Conservative **Pros:** - already implemented - will never miss events - no overhead **Cons:** - tons of repeated work - doesn't allow users to avoid repeating work (or monitoring for other changes) ## Proposal 1: Earlier-This-Tick Change Detection The current approach as of Bevy 0.4. Flags are set, and then flushed at the end of each frame. **Type:** Flaky **Pros:** - already implemented - simple to understand - low memory overhead (2 bits per component) - low time overhead (clear every flag once per frame) **Cons:** - misses systems based on ordering - systems that don't run every frame miss changes - duplicates detection when looping - can lead to unresolvable circular dependencies ## Proposal 2: Two-Tick Change Detection Flags persist for two frames, using a double-buffer system identical to that used in events. A change is observed if it is found in either the current frame's list of changes or the previous frame's. **Type:** Conservative **Pros:** - easy to understand - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - can result in a great deal of duplicated work - systems that don't run every frame miss changes - duplicates detection when looping ## Proposal 3: Last-Tick Change Detection Flags persist for two frames, using a double-buffer system identical to that used in events. A change is observed if it is found in the previous frame's list of changes. **Type:** Exact **Pros:** - exact - easy to understand - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - change detection is always delayed, possibly causing painful chained delays - systems that don't run every frame miss changes - duplicates detection when looping ## Proposal 4: Flag-Doubling Change Detection Combine Proposal 2 and Proposal 3. Differentiate between `JustChanged` (current behavior) and `Changed` (Proposal 3). Pack this data into the flags according to [this implementation proposal](https://github.com/bevyengine/bevy/issues/68#issuecomment-769174804). **Type:** Flaky + Exact **Pros:** - allows users to acc - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - users must specify the type of change detection required - still quite fragile to system ordering effects when using the flaky `JustChanged` form - cannot get immediate + exact results - systems that don't run every frame miss changes - duplicates detection when looping ## [SELECTED] Proposal 5: Generation-Counter Change Detection A global counter is increased after each system is run. Each component saves the time of last mutation, and each system saves the time of last execution. Mutation is detected when the component's counter is greater than the system's counter. Discussed [here](https://github.com/bevyengine/bevy/issues/68#issuecomment-769174804). How to handle addition detection is unsolved; the current proposal is to use the highest bit of the counter as in proposal 1. **Type:** Exact (for mutations), flaky (for additions) **Pros:** - low time overhead (set component counter on access, set system counter after execution) - robust to systems that don't run every frame - robust to systems that loop **Cons:** - moderately complex implementation - must be modified as systems are inserted dynamically - medium memory overhead (4 bytes per component + system) - unsolved addition detection ## Proposal 6: System-Data Change Detection For each system, track which system's changes it has seen. This approach is only worth fully designing and implementing if Proposal 5 fails in some way. **Type:** Exact **Pros:** - exact - conceptually simple **Cons:** - requires storing data on each system - implementation is complex - must be modified as systems are inserted dynamically ## Proposal 7: Total-Order Change Detection Discussed [here](https://github.com/bevyengine/bevy/issues/68#issuecomment-754326523). This proposal is somewhat complicated by the new scheduler, but I believe it should still be conceptually feasible. This approach is only worth fully designing and implementing if Proposal 5 fails in some way. **Type:** Exact **Pros:** - exact - efficient data storage relative to other exact proposals **Cons:** - requires access to the scheduler - complex implementation and difficulty grokking - must be modified as systems are inserted dynamically # Tests - We will need to verify properties 1, 2, 3, 7 and 8. Priority: 1 > 2 = 3 > 8 > 7 - Ideally we can use identical user-facing syntax for all proposals, allowing us to re-use the same syntax for each. - When writing tests, we need to carefully specify order using explicit dependencies. - These tests will need to be duplicated for both components and resources. - We need to be sure to handle cases where ambiguous system orders exist. `changing_system` is always the system that makes the changes, and `detecting_system` always detects the changes. The component / resource changed will be simple boolean wrapper structs. ## Basic Added / Mutated / Changed 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs before `detecting_system` - verify at the end of tick 2 ## At Least Once 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs after `detecting_system` - verify at the end of tick 2 ## At Most Once 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs once before `detecting_system` - increment a counter based on the number of changes detected - verify at the end of tick 2 ## Fast Detection 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs before `detecting_system` - verify at the end of tick 1 ## Ambiguous System Ordering Robustness 2 x 3 x 2 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs [before/after] `detecting_system` in tick 1 - `changing_system` runs [after/before] `detecting_system` in tick 2 ## System Pausing 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs in tick 1, then is disabled by run criteria - `detecting_system` is disabled by run criteria until it is run once during tick 3 - verify at the end of tick 3 ## Addition Causes Mutation 2 design: - Resources vs. Components - `adding_system_1` adds a component / resource - `adding system_2` adds the same component / resource - verify the `Mutated` flag at the end of the tick - verify the `Added` flag at the end of the tick First check tests for: https://github.com/bevyengine/bevy/issues/333 Second check tests for: https://github.com/bevyengine/bevy/issues/1443 ## Changes Made By Commands - `adding_system` runs in Update in tick 1, and sends a command to add a component - `detecting_system` runs in Update in tick 1 and 2, after `adding_system` - We can't detect the changes in tick 1, since they haven't been processed yet - If we were to track these changes as being emitted by `adding_system`, we can't detect the changes in tick 2 either, since `detecting_system` has already run once after `adding_system` :( # Benchmarks See: [general advice](https://github.com/bevyengine/bevy/blob/master/docs/profiling.md), [Criterion crate](https://github.com/bheisler/criterion.rs) There are several critical parameters to vary: 1. entity count (1 to 10^9) 2. fraction of entities that are changed (0% to 100%) 3. cost to perform work on changed entities, i.e. workload (1 ns to 1s) 1 and 2 should be varied between benchmark runs. 3 can be added on computationally. We want to measure: - memory cost - run time We should collect these measurements across several frames (100?) to reduce bootup effects and accurately measure the mean, variance and drift. Entity-component change detection is much more important to benchmark than resource change detection, due to the orders of magnitude higher number of pieces of data. No change detection at all should be included in benchmarks as a second control for cases where missing changes is unacceptable. ## Graphs 1. y: performance, x: log_10(entity count), color: proposal, facet: performance metric. Set cost to perform work to 0. 2. y: run time, x: cost to perform work, color: proposal, facet: fraction changed. Set number of entities to 10^6 3. y: memory, x: frames, color: proposal # Conclusions 1. Is the theoretical categorization of the proposals correct according to our tests? 2. How does the performance of the proposals compare without any load? 3. How does the performance of the proposals compare with realistic loads? 4. At what workload does more exact change tracking become worth the (presumably) higher overhead? 5. When does adding change-detection to save on work become worthwhile? 6. Is there enough divergence in performance between the best solutions in each class to ship more than one change-tracking solution? # Implementation Plan 1. Write a test suite. 2. Verify that tests fail for existing approach. 3. Write a benchmark suite. 4. Get performance numbers for existing approach. 5. Implement, test and benchmark various solutions using a Git branch per proposal. 6. Create a draft PR with all solutions and present results to team. 7. Select a solution and replace existing change detection. Co-authored-by: Brice DAVIER <bricedavier@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2021-03-19 17:53:26 +00:00
get_filtered::<Changed<A>>(&mut world),
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
vec![e3],
"e1 should no longer be returned"
);
world.clear_trackers();
Reliable change detection (#1471) # Problem Definition The current change tracking (via flags for both components and resources) fails to detect changes made by systems that are scheduled to run earlier in the frame than they are. This issue is discussed at length in [#68](https://github.com/bevyengine/bevy/issues/68) and [#54](https://github.com/bevyengine/bevy/issues/54). This is very much a draft PR, and contributions are welcome and needed. # Criteria 1. Each change is detected at least once, no matter the ordering. 2. Each change is detected at most once, no matter the ordering. 3. Changes should be detected the same frame that they are made. 4. Competitive ergonomics. Ideally does not require opting-in. 5. Low CPU overhead of computation. 6. Memory efficient. This must not increase over time, except where the number of entities / resources does. 7. Changes should not be lost for systems that don't run. 8. A frame needs to act as a pure function. Given the same set of entities / components it needs to produce the same end state without side-effects. **Exact** change-tracking proposals satisfy criteria 1 and 2. **Conservative** change-tracking proposals satisfy criteria 1 but not 2. **Flaky** change tracking proposals satisfy criteria 2 but not 1. # Code Base Navigation There are three types of flags: - `Added`: A piece of data was added to an entity / `Resources`. - `Mutated`: A piece of data was able to be modified, because its `DerefMut` was accessed - `Changed`: The bitwise OR of `Added` and `Changed` The special behavior of `ChangedRes`, with respect to the scheduler is being removed in [#1313](https://github.com/bevyengine/bevy/pull/1313) and does not need to be reproduced. `ChangedRes` and friends can be found in "bevy_ecs/core/resources/resource_query.rs". The `Flags` trait for Components can be found in "bevy_ecs/core/query.rs". `ComponentFlags` are stored in "bevy_ecs/core/archetypes.rs", defined on line 446. # Proposals **Proposal 5 was selected for implementation.** ## Proposal 0: No Change Detection The baseline, where computations are performed on everything regardless of whether it changed. **Type:** Conservative **Pros:** - already implemented - will never miss events - no overhead **Cons:** - tons of repeated work - doesn't allow users to avoid repeating work (or monitoring for other changes) ## Proposal 1: Earlier-This-Tick Change Detection The current approach as of Bevy 0.4. Flags are set, and then flushed at the end of each frame. **Type:** Flaky **Pros:** - already implemented - simple to understand - low memory overhead (2 bits per component) - low time overhead (clear every flag once per frame) **Cons:** - misses systems based on ordering - systems that don't run every frame miss changes - duplicates detection when looping - can lead to unresolvable circular dependencies ## Proposal 2: Two-Tick Change Detection Flags persist for two frames, using a double-buffer system identical to that used in events. A change is observed if it is found in either the current frame's list of changes or the previous frame's. **Type:** Conservative **Pros:** - easy to understand - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - can result in a great deal of duplicated work - systems that don't run every frame miss changes - duplicates detection when looping ## Proposal 3: Last-Tick Change Detection Flags persist for two frames, using a double-buffer system identical to that used in events. A change is observed if it is found in the previous frame's list of changes. **Type:** Exact **Pros:** - exact - easy to understand - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - change detection is always delayed, possibly causing painful chained delays - systems that don't run every frame miss changes - duplicates detection when looping ## Proposal 4: Flag-Doubling Change Detection Combine Proposal 2 and Proposal 3. Differentiate between `JustChanged` (current behavior) and `Changed` (Proposal 3). Pack this data into the flags according to [this implementation proposal](https://github.com/bevyengine/bevy/issues/68#issuecomment-769174804). **Type:** Flaky + Exact **Pros:** - allows users to acc - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - users must specify the type of change detection required - still quite fragile to system ordering effects when using the flaky `JustChanged` form - cannot get immediate + exact results - systems that don't run every frame miss changes - duplicates detection when looping ## [SELECTED] Proposal 5: Generation-Counter Change Detection A global counter is increased after each system is run. Each component saves the time of last mutation, and each system saves the time of last execution. Mutation is detected when the component's counter is greater than the system's counter. Discussed [here](https://github.com/bevyengine/bevy/issues/68#issuecomment-769174804). How to handle addition detection is unsolved; the current proposal is to use the highest bit of the counter as in proposal 1. **Type:** Exact (for mutations), flaky (for additions) **Pros:** - low time overhead (set component counter on access, set system counter after execution) - robust to systems that don't run every frame - robust to systems that loop **Cons:** - moderately complex implementation - must be modified as systems are inserted dynamically - medium memory overhead (4 bytes per component + system) - unsolved addition detection ## Proposal 6: System-Data Change Detection For each system, track which system's changes it has seen. This approach is only worth fully designing and implementing if Proposal 5 fails in some way. **Type:** Exact **Pros:** - exact - conceptually simple **Cons:** - requires storing data on each system - implementation is complex - must be modified as systems are inserted dynamically ## Proposal 7: Total-Order Change Detection Discussed [here](https://github.com/bevyengine/bevy/issues/68#issuecomment-754326523). This proposal is somewhat complicated by the new scheduler, but I believe it should still be conceptually feasible. This approach is only worth fully designing and implementing if Proposal 5 fails in some way. **Type:** Exact **Pros:** - exact - efficient data storage relative to other exact proposals **Cons:** - requires access to the scheduler - complex implementation and difficulty grokking - must be modified as systems are inserted dynamically # Tests - We will need to verify properties 1, 2, 3, 7 and 8. Priority: 1 > 2 = 3 > 8 > 7 - Ideally we can use identical user-facing syntax for all proposals, allowing us to re-use the same syntax for each. - When writing tests, we need to carefully specify order using explicit dependencies. - These tests will need to be duplicated for both components and resources. - We need to be sure to handle cases where ambiguous system orders exist. `changing_system` is always the system that makes the changes, and `detecting_system` always detects the changes. The component / resource changed will be simple boolean wrapper structs. ## Basic Added / Mutated / Changed 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs before `detecting_system` - verify at the end of tick 2 ## At Least Once 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs after `detecting_system` - verify at the end of tick 2 ## At Most Once 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs once before `detecting_system` - increment a counter based on the number of changes detected - verify at the end of tick 2 ## Fast Detection 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs before `detecting_system` - verify at the end of tick 1 ## Ambiguous System Ordering Robustness 2 x 3 x 2 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs [before/after] `detecting_system` in tick 1 - `changing_system` runs [after/before] `detecting_system` in tick 2 ## System Pausing 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs in tick 1, then is disabled by run criteria - `detecting_system` is disabled by run criteria until it is run once during tick 3 - verify at the end of tick 3 ## Addition Causes Mutation 2 design: - Resources vs. Components - `adding_system_1` adds a component / resource - `adding system_2` adds the same component / resource - verify the `Mutated` flag at the end of the tick - verify the `Added` flag at the end of the tick First check tests for: https://github.com/bevyengine/bevy/issues/333 Second check tests for: https://github.com/bevyengine/bevy/issues/1443 ## Changes Made By Commands - `adding_system` runs in Update in tick 1, and sends a command to add a component - `detecting_system` runs in Update in tick 1 and 2, after `adding_system` - We can't detect the changes in tick 1, since they haven't been processed yet - If we were to track these changes as being emitted by `adding_system`, we can't detect the changes in tick 2 either, since `detecting_system` has already run once after `adding_system` :( # Benchmarks See: [general advice](https://github.com/bevyengine/bevy/blob/master/docs/profiling.md), [Criterion crate](https://github.com/bheisler/criterion.rs) There are several critical parameters to vary: 1. entity count (1 to 10^9) 2. fraction of entities that are changed (0% to 100%) 3. cost to perform work on changed entities, i.e. workload (1 ns to 1s) 1 and 2 should be varied between benchmark runs. 3 can be added on computationally. We want to measure: - memory cost - run time We should collect these measurements across several frames (100?) to reduce bootup effects and accurately measure the mean, variance and drift. Entity-component change detection is much more important to benchmark than resource change detection, due to the orders of magnitude higher number of pieces of data. No change detection at all should be included in benchmarks as a second control for cases where missing changes is unacceptable. ## Graphs 1. y: performance, x: log_10(entity count), color: proposal, facet: performance metric. Set cost to perform work to 0. 2. y: run time, x: cost to perform work, color: proposal, facet: fraction changed. Set number of entities to 10^6 3. y: memory, x: frames, color: proposal # Conclusions 1. Is the theoretical categorization of the proposals correct according to our tests? 2. How does the performance of the proposals compare without any load? 3. How does the performance of the proposals compare with realistic loads? 4. At what workload does more exact change tracking become worth the (presumably) higher overhead? 5. When does adding change-detection to save on work become worthwhile? 6. Is there enough divergence in performance between the best solutions in each class to ship more than one change-tracking solution? # Implementation Plan 1. Write a test suite. 2. Verify that tests fail for existing approach. 3. Write a benchmark suite. 4. Get performance numbers for existing approach. 5. Implement, test and benchmark various solutions using a Git branch per proposal. 6. Create a draft PR with all solutions and present results to team. 7. Select a solution and replace existing change detection. Co-authored-by: Brice DAVIER <bricedavier@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2021-03-19 17:53:26 +00:00
assert!(get_filtered::<Changed<A>>(&mut world).is_empty());
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
let e4 = world.spawn_empty().id();
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
world.entity_mut(e4).insert(A(0));
Reliable change detection (#1471) # Problem Definition The current change tracking (via flags for both components and resources) fails to detect changes made by systems that are scheduled to run earlier in the frame than they are. This issue is discussed at length in [#68](https://github.com/bevyengine/bevy/issues/68) and [#54](https://github.com/bevyengine/bevy/issues/54). This is very much a draft PR, and contributions are welcome and needed. # Criteria 1. Each change is detected at least once, no matter the ordering. 2. Each change is detected at most once, no matter the ordering. 3. Changes should be detected the same frame that they are made. 4. Competitive ergonomics. Ideally does not require opting-in. 5. Low CPU overhead of computation. 6. Memory efficient. This must not increase over time, except where the number of entities / resources does. 7. Changes should not be lost for systems that don't run. 8. A frame needs to act as a pure function. Given the same set of entities / components it needs to produce the same end state without side-effects. **Exact** change-tracking proposals satisfy criteria 1 and 2. **Conservative** change-tracking proposals satisfy criteria 1 but not 2. **Flaky** change tracking proposals satisfy criteria 2 but not 1. # Code Base Navigation There are three types of flags: - `Added`: A piece of data was added to an entity / `Resources`. - `Mutated`: A piece of data was able to be modified, because its `DerefMut` was accessed - `Changed`: The bitwise OR of `Added` and `Changed` The special behavior of `ChangedRes`, with respect to the scheduler is being removed in [#1313](https://github.com/bevyengine/bevy/pull/1313) and does not need to be reproduced. `ChangedRes` and friends can be found in "bevy_ecs/core/resources/resource_query.rs". The `Flags` trait for Components can be found in "bevy_ecs/core/query.rs". `ComponentFlags` are stored in "bevy_ecs/core/archetypes.rs", defined on line 446. # Proposals **Proposal 5 was selected for implementation.** ## Proposal 0: No Change Detection The baseline, where computations are performed on everything regardless of whether it changed. **Type:** Conservative **Pros:** - already implemented - will never miss events - no overhead **Cons:** - tons of repeated work - doesn't allow users to avoid repeating work (or monitoring for other changes) ## Proposal 1: Earlier-This-Tick Change Detection The current approach as of Bevy 0.4. Flags are set, and then flushed at the end of each frame. **Type:** Flaky **Pros:** - already implemented - simple to understand - low memory overhead (2 bits per component) - low time overhead (clear every flag once per frame) **Cons:** - misses systems based on ordering - systems that don't run every frame miss changes - duplicates detection when looping - can lead to unresolvable circular dependencies ## Proposal 2: Two-Tick Change Detection Flags persist for two frames, using a double-buffer system identical to that used in events. A change is observed if it is found in either the current frame's list of changes or the previous frame's. **Type:** Conservative **Pros:** - easy to understand - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - can result in a great deal of duplicated work - systems that don't run every frame miss changes - duplicates detection when looping ## Proposal 3: Last-Tick Change Detection Flags persist for two frames, using a double-buffer system identical to that used in events. A change is observed if it is found in the previous frame's list of changes. **Type:** Exact **Pros:** - exact - easy to understand - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - change detection is always delayed, possibly causing painful chained delays - systems that don't run every frame miss changes - duplicates detection when looping ## Proposal 4: Flag-Doubling Change Detection Combine Proposal 2 and Proposal 3. Differentiate between `JustChanged` (current behavior) and `Changed` (Proposal 3). Pack this data into the flags according to [this implementation proposal](https://github.com/bevyengine/bevy/issues/68#issuecomment-769174804). **Type:** Flaky + Exact **Pros:** - allows users to acc - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - users must specify the type of change detection required - still quite fragile to system ordering effects when using the flaky `JustChanged` form - cannot get immediate + exact results - systems that don't run every frame miss changes - duplicates detection when looping ## [SELECTED] Proposal 5: Generation-Counter Change Detection A global counter is increased after each system is run. Each component saves the time of last mutation, and each system saves the time of last execution. Mutation is detected when the component's counter is greater than the system's counter. Discussed [here](https://github.com/bevyengine/bevy/issues/68#issuecomment-769174804). How to handle addition detection is unsolved; the current proposal is to use the highest bit of the counter as in proposal 1. **Type:** Exact (for mutations), flaky (for additions) **Pros:** - low time overhead (set component counter on access, set system counter after execution) - robust to systems that don't run every frame - robust to systems that loop **Cons:** - moderately complex implementation - must be modified as systems are inserted dynamically - medium memory overhead (4 bytes per component + system) - unsolved addition detection ## Proposal 6: System-Data Change Detection For each system, track which system's changes it has seen. This approach is only worth fully designing and implementing if Proposal 5 fails in some way. **Type:** Exact **Pros:** - exact - conceptually simple **Cons:** - requires storing data on each system - implementation is complex - must be modified as systems are inserted dynamically ## Proposal 7: Total-Order Change Detection Discussed [here](https://github.com/bevyengine/bevy/issues/68#issuecomment-754326523). This proposal is somewhat complicated by the new scheduler, but I believe it should still be conceptually feasible. This approach is only worth fully designing and implementing if Proposal 5 fails in some way. **Type:** Exact **Pros:** - exact - efficient data storage relative to other exact proposals **Cons:** - requires access to the scheduler - complex implementation and difficulty grokking - must be modified as systems are inserted dynamically # Tests - We will need to verify properties 1, 2, 3, 7 and 8. Priority: 1 > 2 = 3 > 8 > 7 - Ideally we can use identical user-facing syntax for all proposals, allowing us to re-use the same syntax for each. - When writing tests, we need to carefully specify order using explicit dependencies. - These tests will need to be duplicated for both components and resources. - We need to be sure to handle cases where ambiguous system orders exist. `changing_system` is always the system that makes the changes, and `detecting_system` always detects the changes. The component / resource changed will be simple boolean wrapper structs. ## Basic Added / Mutated / Changed 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs before `detecting_system` - verify at the end of tick 2 ## At Least Once 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs after `detecting_system` - verify at the end of tick 2 ## At Most Once 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs once before `detecting_system` - increment a counter based on the number of changes detected - verify at the end of tick 2 ## Fast Detection 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs before `detecting_system` - verify at the end of tick 1 ## Ambiguous System Ordering Robustness 2 x 3 x 2 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs [before/after] `detecting_system` in tick 1 - `changing_system` runs [after/before] `detecting_system` in tick 2 ## System Pausing 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs in tick 1, then is disabled by run criteria - `detecting_system` is disabled by run criteria until it is run once during tick 3 - verify at the end of tick 3 ## Addition Causes Mutation 2 design: - Resources vs. Components - `adding_system_1` adds a component / resource - `adding system_2` adds the same component / resource - verify the `Mutated` flag at the end of the tick - verify the `Added` flag at the end of the tick First check tests for: https://github.com/bevyengine/bevy/issues/333 Second check tests for: https://github.com/bevyengine/bevy/issues/1443 ## Changes Made By Commands - `adding_system` runs in Update in tick 1, and sends a command to add a component - `detecting_system` runs in Update in tick 1 and 2, after `adding_system` - We can't detect the changes in tick 1, since they haven't been processed yet - If we were to track these changes as being emitted by `adding_system`, we can't detect the changes in tick 2 either, since `detecting_system` has already run once after `adding_system` :( # Benchmarks See: [general advice](https://github.com/bevyengine/bevy/blob/master/docs/profiling.md), [Criterion crate](https://github.com/bheisler/criterion.rs) There are several critical parameters to vary: 1. entity count (1 to 10^9) 2. fraction of entities that are changed (0% to 100%) 3. cost to perform work on changed entities, i.e. workload (1 ns to 1s) 1 and 2 should be varied between benchmark runs. 3 can be added on computationally. We want to measure: - memory cost - run time We should collect these measurements across several frames (100?) to reduce bootup effects and accurately measure the mean, variance and drift. Entity-component change detection is much more important to benchmark than resource change detection, due to the orders of magnitude higher number of pieces of data. No change detection at all should be included in benchmarks as a second control for cases where missing changes is unacceptable. ## Graphs 1. y: performance, x: log_10(entity count), color: proposal, facet: performance metric. Set cost to perform work to 0. 2. y: run time, x: cost to perform work, color: proposal, facet: fraction changed. Set number of entities to 10^6 3. y: memory, x: frames, color: proposal # Conclusions 1. Is the theoretical categorization of the proposals correct according to our tests? 2. How does the performance of the proposals compare without any load? 3. How does the performance of the proposals compare with realistic loads? 4. At what workload does more exact change tracking become worth the (presumably) higher overhead? 5. When does adding change-detection to save on work become worthwhile? 6. Is there enough divergence in performance between the best solutions in each class to ship more than one change-tracking solution? # Implementation Plan 1. Write a test suite. 2. Verify that tests fail for existing approach. 3. Write a benchmark suite. 4. Get performance numbers for existing approach. 5. Implement, test and benchmark various solutions using a Git branch per proposal. 6. Create a draft PR with all solutions and present results to team. 7. Select a solution and replace existing change detection. Co-authored-by: Brice DAVIER <bricedavier@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2021-03-19 17:53:26 +00:00
assert_eq!(get_filtered::<Changed<A>>(&mut world), vec![e4]);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
assert_eq!(get_filtered::<Added<A>>(&mut world), vec![e4]);
world.entity_mut(e4).insert(A(1));
Reliable change detection (#1471) # Problem Definition The current change tracking (via flags for both components and resources) fails to detect changes made by systems that are scheduled to run earlier in the frame than they are. This issue is discussed at length in [#68](https://github.com/bevyengine/bevy/issues/68) and [#54](https://github.com/bevyengine/bevy/issues/54). This is very much a draft PR, and contributions are welcome and needed. # Criteria 1. Each change is detected at least once, no matter the ordering. 2. Each change is detected at most once, no matter the ordering. 3. Changes should be detected the same frame that they are made. 4. Competitive ergonomics. Ideally does not require opting-in. 5. Low CPU overhead of computation. 6. Memory efficient. This must not increase over time, except where the number of entities / resources does. 7. Changes should not be lost for systems that don't run. 8. A frame needs to act as a pure function. Given the same set of entities / components it needs to produce the same end state without side-effects. **Exact** change-tracking proposals satisfy criteria 1 and 2. **Conservative** change-tracking proposals satisfy criteria 1 but not 2. **Flaky** change tracking proposals satisfy criteria 2 but not 1. # Code Base Navigation There are three types of flags: - `Added`: A piece of data was added to an entity / `Resources`. - `Mutated`: A piece of data was able to be modified, because its `DerefMut` was accessed - `Changed`: The bitwise OR of `Added` and `Changed` The special behavior of `ChangedRes`, with respect to the scheduler is being removed in [#1313](https://github.com/bevyengine/bevy/pull/1313) and does not need to be reproduced. `ChangedRes` and friends can be found in "bevy_ecs/core/resources/resource_query.rs". The `Flags` trait for Components can be found in "bevy_ecs/core/query.rs". `ComponentFlags` are stored in "bevy_ecs/core/archetypes.rs", defined on line 446. # Proposals **Proposal 5 was selected for implementation.** ## Proposal 0: No Change Detection The baseline, where computations are performed on everything regardless of whether it changed. **Type:** Conservative **Pros:** - already implemented - will never miss events - no overhead **Cons:** - tons of repeated work - doesn't allow users to avoid repeating work (or monitoring for other changes) ## Proposal 1: Earlier-This-Tick Change Detection The current approach as of Bevy 0.4. Flags are set, and then flushed at the end of each frame. **Type:** Flaky **Pros:** - already implemented - simple to understand - low memory overhead (2 bits per component) - low time overhead (clear every flag once per frame) **Cons:** - misses systems based on ordering - systems that don't run every frame miss changes - duplicates detection when looping - can lead to unresolvable circular dependencies ## Proposal 2: Two-Tick Change Detection Flags persist for two frames, using a double-buffer system identical to that used in events. A change is observed if it is found in either the current frame's list of changes or the previous frame's. **Type:** Conservative **Pros:** - easy to understand - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - can result in a great deal of duplicated work - systems that don't run every frame miss changes - duplicates detection when looping ## Proposal 3: Last-Tick Change Detection Flags persist for two frames, using a double-buffer system identical to that used in events. A change is observed if it is found in the previous frame's list of changes. **Type:** Exact **Pros:** - exact - easy to understand - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - change detection is always delayed, possibly causing painful chained delays - systems that don't run every frame miss changes - duplicates detection when looping ## Proposal 4: Flag-Doubling Change Detection Combine Proposal 2 and Proposal 3. Differentiate between `JustChanged` (current behavior) and `Changed` (Proposal 3). Pack this data into the flags according to [this implementation proposal](https://github.com/bevyengine/bevy/issues/68#issuecomment-769174804). **Type:** Flaky + Exact **Pros:** - allows users to acc - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - users must specify the type of change detection required - still quite fragile to system ordering effects when using the flaky `JustChanged` form - cannot get immediate + exact results - systems that don't run every frame miss changes - duplicates detection when looping ## [SELECTED] Proposal 5: Generation-Counter Change Detection A global counter is increased after each system is run. Each component saves the time of last mutation, and each system saves the time of last execution. Mutation is detected when the component's counter is greater than the system's counter. Discussed [here](https://github.com/bevyengine/bevy/issues/68#issuecomment-769174804). How to handle addition detection is unsolved; the current proposal is to use the highest bit of the counter as in proposal 1. **Type:** Exact (for mutations), flaky (for additions) **Pros:** - low time overhead (set component counter on access, set system counter after execution) - robust to systems that don't run every frame - robust to systems that loop **Cons:** - moderately complex implementation - must be modified as systems are inserted dynamically - medium memory overhead (4 bytes per component + system) - unsolved addition detection ## Proposal 6: System-Data Change Detection For each system, track which system's changes it has seen. This approach is only worth fully designing and implementing if Proposal 5 fails in some way. **Type:** Exact **Pros:** - exact - conceptually simple **Cons:** - requires storing data on each system - implementation is complex - must be modified as systems are inserted dynamically ## Proposal 7: Total-Order Change Detection Discussed [here](https://github.com/bevyengine/bevy/issues/68#issuecomment-754326523). This proposal is somewhat complicated by the new scheduler, but I believe it should still be conceptually feasible. This approach is only worth fully designing and implementing if Proposal 5 fails in some way. **Type:** Exact **Pros:** - exact - efficient data storage relative to other exact proposals **Cons:** - requires access to the scheduler - complex implementation and difficulty grokking - must be modified as systems are inserted dynamically # Tests - We will need to verify properties 1, 2, 3, 7 and 8. Priority: 1 > 2 = 3 > 8 > 7 - Ideally we can use identical user-facing syntax for all proposals, allowing us to re-use the same syntax for each. - When writing tests, we need to carefully specify order using explicit dependencies. - These tests will need to be duplicated for both components and resources. - We need to be sure to handle cases where ambiguous system orders exist. `changing_system` is always the system that makes the changes, and `detecting_system` always detects the changes. The component / resource changed will be simple boolean wrapper structs. ## Basic Added / Mutated / Changed 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs before `detecting_system` - verify at the end of tick 2 ## At Least Once 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs after `detecting_system` - verify at the end of tick 2 ## At Most Once 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs once before `detecting_system` - increment a counter based on the number of changes detected - verify at the end of tick 2 ## Fast Detection 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs before `detecting_system` - verify at the end of tick 1 ## Ambiguous System Ordering Robustness 2 x 3 x 2 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs [before/after] `detecting_system` in tick 1 - `changing_system` runs [after/before] `detecting_system` in tick 2 ## System Pausing 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs in tick 1, then is disabled by run criteria - `detecting_system` is disabled by run criteria until it is run once during tick 3 - verify at the end of tick 3 ## Addition Causes Mutation 2 design: - Resources vs. Components - `adding_system_1` adds a component / resource - `adding system_2` adds the same component / resource - verify the `Mutated` flag at the end of the tick - verify the `Added` flag at the end of the tick First check tests for: https://github.com/bevyengine/bevy/issues/333 Second check tests for: https://github.com/bevyengine/bevy/issues/1443 ## Changes Made By Commands - `adding_system` runs in Update in tick 1, and sends a command to add a component - `detecting_system` runs in Update in tick 1 and 2, after `adding_system` - We can't detect the changes in tick 1, since they haven't been processed yet - If we were to track these changes as being emitted by `adding_system`, we can't detect the changes in tick 2 either, since `detecting_system` has already run once after `adding_system` :( # Benchmarks See: [general advice](https://github.com/bevyengine/bevy/blob/master/docs/profiling.md), [Criterion crate](https://github.com/bheisler/criterion.rs) There are several critical parameters to vary: 1. entity count (1 to 10^9) 2. fraction of entities that are changed (0% to 100%) 3. cost to perform work on changed entities, i.e. workload (1 ns to 1s) 1 and 2 should be varied between benchmark runs. 3 can be added on computationally. We want to measure: - memory cost - run time We should collect these measurements across several frames (100?) to reduce bootup effects and accurately measure the mean, variance and drift. Entity-component change detection is much more important to benchmark than resource change detection, due to the orders of magnitude higher number of pieces of data. No change detection at all should be included in benchmarks as a second control for cases where missing changes is unacceptable. ## Graphs 1. y: performance, x: log_10(entity count), color: proposal, facet: performance metric. Set cost to perform work to 0. 2. y: run time, x: cost to perform work, color: proposal, facet: fraction changed. Set number of entities to 10^6 3. y: memory, x: frames, color: proposal # Conclusions 1. Is the theoretical categorization of the proposals correct according to our tests? 2. How does the performance of the proposals compare without any load? 3. How does the performance of the proposals compare with realistic loads? 4. At what workload does more exact change tracking become worth the (presumably) higher overhead? 5. When does adding change-detection to save on work become worthwhile? 6. Is there enough divergence in performance between the best solutions in each class to ship more than one change-tracking solution? # Implementation Plan 1. Write a test suite. 2. Verify that tests fail for existing approach. 3. Write a benchmark suite. 4. Get performance numbers for existing approach. 5. Implement, test and benchmark various solutions using a Git branch per proposal. 6. Create a draft PR with all solutions and present results to team. 7. Select a solution and replace existing change detection. Co-authored-by: Brice DAVIER <bricedavier@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2021-03-19 17:53:26 +00:00
assert_eq!(get_filtered::<Changed<A>>(&mut world), vec![e4]);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
world.clear_trackers();
Reliable change detection (#1471) # Problem Definition The current change tracking (via flags for both components and resources) fails to detect changes made by systems that are scheduled to run earlier in the frame than they are. This issue is discussed at length in [#68](https://github.com/bevyengine/bevy/issues/68) and [#54](https://github.com/bevyengine/bevy/issues/54). This is very much a draft PR, and contributions are welcome and needed. # Criteria 1. Each change is detected at least once, no matter the ordering. 2. Each change is detected at most once, no matter the ordering. 3. Changes should be detected the same frame that they are made. 4. Competitive ergonomics. Ideally does not require opting-in. 5. Low CPU overhead of computation. 6. Memory efficient. This must not increase over time, except where the number of entities / resources does. 7. Changes should not be lost for systems that don't run. 8. A frame needs to act as a pure function. Given the same set of entities / components it needs to produce the same end state without side-effects. **Exact** change-tracking proposals satisfy criteria 1 and 2. **Conservative** change-tracking proposals satisfy criteria 1 but not 2. **Flaky** change tracking proposals satisfy criteria 2 but not 1. # Code Base Navigation There are three types of flags: - `Added`: A piece of data was added to an entity / `Resources`. - `Mutated`: A piece of data was able to be modified, because its `DerefMut` was accessed - `Changed`: The bitwise OR of `Added` and `Changed` The special behavior of `ChangedRes`, with respect to the scheduler is being removed in [#1313](https://github.com/bevyengine/bevy/pull/1313) and does not need to be reproduced. `ChangedRes` and friends can be found in "bevy_ecs/core/resources/resource_query.rs". The `Flags` trait for Components can be found in "bevy_ecs/core/query.rs". `ComponentFlags` are stored in "bevy_ecs/core/archetypes.rs", defined on line 446. # Proposals **Proposal 5 was selected for implementation.** ## Proposal 0: No Change Detection The baseline, where computations are performed on everything regardless of whether it changed. **Type:** Conservative **Pros:** - already implemented - will never miss events - no overhead **Cons:** - tons of repeated work - doesn't allow users to avoid repeating work (or monitoring for other changes) ## Proposal 1: Earlier-This-Tick Change Detection The current approach as of Bevy 0.4. Flags are set, and then flushed at the end of each frame. **Type:** Flaky **Pros:** - already implemented - simple to understand - low memory overhead (2 bits per component) - low time overhead (clear every flag once per frame) **Cons:** - misses systems based on ordering - systems that don't run every frame miss changes - duplicates detection when looping - can lead to unresolvable circular dependencies ## Proposal 2: Two-Tick Change Detection Flags persist for two frames, using a double-buffer system identical to that used in events. A change is observed if it is found in either the current frame's list of changes or the previous frame's. **Type:** Conservative **Pros:** - easy to understand - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - can result in a great deal of duplicated work - systems that don't run every frame miss changes - duplicates detection when looping ## Proposal 3: Last-Tick Change Detection Flags persist for two frames, using a double-buffer system identical to that used in events. A change is observed if it is found in the previous frame's list of changes. **Type:** Exact **Pros:** - exact - easy to understand - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - change detection is always delayed, possibly causing painful chained delays - systems that don't run every frame miss changes - duplicates detection when looping ## Proposal 4: Flag-Doubling Change Detection Combine Proposal 2 and Proposal 3. Differentiate between `JustChanged` (current behavior) and `Changed` (Proposal 3). Pack this data into the flags according to [this implementation proposal](https://github.com/bevyengine/bevy/issues/68#issuecomment-769174804). **Type:** Flaky + Exact **Pros:** - allows users to acc - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - users must specify the type of change detection required - still quite fragile to system ordering effects when using the flaky `JustChanged` form - cannot get immediate + exact results - systems that don't run every frame miss changes - duplicates detection when looping ## [SELECTED] Proposal 5: Generation-Counter Change Detection A global counter is increased after each system is run. Each component saves the time of last mutation, and each system saves the time of last execution. Mutation is detected when the component's counter is greater than the system's counter. Discussed [here](https://github.com/bevyengine/bevy/issues/68#issuecomment-769174804). How to handle addition detection is unsolved; the current proposal is to use the highest bit of the counter as in proposal 1. **Type:** Exact (for mutations), flaky (for additions) **Pros:** - low time overhead (set component counter on access, set system counter after execution) - robust to systems that don't run every frame - robust to systems that loop **Cons:** - moderately complex implementation - must be modified as systems are inserted dynamically - medium memory overhead (4 bytes per component + system) - unsolved addition detection ## Proposal 6: System-Data Change Detection For each system, track which system's changes it has seen. This approach is only worth fully designing and implementing if Proposal 5 fails in some way. **Type:** Exact **Pros:** - exact - conceptually simple **Cons:** - requires storing data on each system - implementation is complex - must be modified as systems are inserted dynamically ## Proposal 7: Total-Order Change Detection Discussed [here](https://github.com/bevyengine/bevy/issues/68#issuecomment-754326523). This proposal is somewhat complicated by the new scheduler, but I believe it should still be conceptually feasible. This approach is only worth fully designing and implementing if Proposal 5 fails in some way. **Type:** Exact **Pros:** - exact - efficient data storage relative to other exact proposals **Cons:** - requires access to the scheduler - complex implementation and difficulty grokking - must be modified as systems are inserted dynamically # Tests - We will need to verify properties 1, 2, 3, 7 and 8. Priority: 1 > 2 = 3 > 8 > 7 - Ideally we can use identical user-facing syntax for all proposals, allowing us to re-use the same syntax for each. - When writing tests, we need to carefully specify order using explicit dependencies. - These tests will need to be duplicated for both components and resources. - We need to be sure to handle cases where ambiguous system orders exist. `changing_system` is always the system that makes the changes, and `detecting_system` always detects the changes. The component / resource changed will be simple boolean wrapper structs. ## Basic Added / Mutated / Changed 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs before `detecting_system` - verify at the end of tick 2 ## At Least Once 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs after `detecting_system` - verify at the end of tick 2 ## At Most Once 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs once before `detecting_system` - increment a counter based on the number of changes detected - verify at the end of tick 2 ## Fast Detection 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs before `detecting_system` - verify at the end of tick 1 ## Ambiguous System Ordering Robustness 2 x 3 x 2 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs [before/after] `detecting_system` in tick 1 - `changing_system` runs [after/before] `detecting_system` in tick 2 ## System Pausing 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs in tick 1, then is disabled by run criteria - `detecting_system` is disabled by run criteria until it is run once during tick 3 - verify at the end of tick 3 ## Addition Causes Mutation 2 design: - Resources vs. Components - `adding_system_1` adds a component / resource - `adding system_2` adds the same component / resource - verify the `Mutated` flag at the end of the tick - verify the `Added` flag at the end of the tick First check tests for: https://github.com/bevyengine/bevy/issues/333 Second check tests for: https://github.com/bevyengine/bevy/issues/1443 ## Changes Made By Commands - `adding_system` runs in Update in tick 1, and sends a command to add a component - `detecting_system` runs in Update in tick 1 and 2, after `adding_system` - We can't detect the changes in tick 1, since they haven't been processed yet - If we were to track these changes as being emitted by `adding_system`, we can't detect the changes in tick 2 either, since `detecting_system` has already run once after `adding_system` :( # Benchmarks See: [general advice](https://github.com/bevyengine/bevy/blob/master/docs/profiling.md), [Criterion crate](https://github.com/bheisler/criterion.rs) There are several critical parameters to vary: 1. entity count (1 to 10^9) 2. fraction of entities that are changed (0% to 100%) 3. cost to perform work on changed entities, i.e. workload (1 ns to 1s) 1 and 2 should be varied between benchmark runs. 3 can be added on computationally. We want to measure: - memory cost - run time We should collect these measurements across several frames (100?) to reduce bootup effects and accurately measure the mean, variance and drift. Entity-component change detection is much more important to benchmark than resource change detection, due to the orders of magnitude higher number of pieces of data. No change detection at all should be included in benchmarks as a second control for cases where missing changes is unacceptable. ## Graphs 1. y: performance, x: log_10(entity count), color: proposal, facet: performance metric. Set cost to perform work to 0. 2. y: run time, x: cost to perform work, color: proposal, facet: fraction changed. Set number of entities to 10^6 3. y: memory, x: frames, color: proposal # Conclusions 1. Is the theoretical categorization of the proposals correct according to our tests? 2. How does the performance of the proposals compare without any load? 3. How does the performance of the proposals compare with realistic loads? 4. At what workload does more exact change tracking become worth the (presumably) higher overhead? 5. When does adding change-detection to save on work become worthwhile? 6. Is there enough divergence in performance between the best solutions in each class to ship more than one change-tracking solution? # Implementation Plan 1. Write a test suite. 2. Verify that tests fail for existing approach. 3. Write a benchmark suite. 4. Get performance numbers for existing approach. 5. Implement, test and benchmark various solutions using a Git branch per proposal. 6. Create a draft PR with all solutions and present results to team. 7. Select a solution and replace existing change detection. Co-authored-by: Brice DAVIER <bricedavier@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2021-03-19 17:53:26 +00:00
// ensure inserting multiple components set changed state for all components and set added
// state for non existing components even when changing archetype.
Accept Bundles for insert and remove. Deprecate insert/remove_bundle (#6039) # Objective Take advantage of the "impl Bundle for Component" changes in #2975 / add the follow up changes discussed there. ## Solution - Change `insert` and `remove` to accept a Bundle instead of a Component (for both Commands and World) - Deprecate `insert_bundle`, `remove_bundle`, and `remove_bundle_intersection` - Add `remove_intersection` --- ## Changelog - Change `insert` and `remove` now accept a Bundle instead of a Component (for both Commands and World) - `insert_bundle` and `remove_bundle` are deprecated ## Migration Guide Replace `insert_bundle` with `insert`: ```rust // Old (0.8) commands.spawn().insert_bundle(SomeBundle::default()); // New (0.9) commands.spawn().insert(SomeBundle::default()); ``` Replace `remove_bundle` with `remove`: ```rust // Old (0.8) commands.entity(some_entity).remove_bundle::<SomeBundle>(); // New (0.9) commands.entity(some_entity).remove::<SomeBundle>(); ``` Replace `remove_bundle_intersection` with `remove_intersection`: ```rust // Old (0.8) world.entity_mut(some_entity).remove_bundle_intersection::<SomeBundle>(); // New (0.9) world.entity_mut(some_entity).remove_intersection::<SomeBundle>(); ``` Consider consolidating as many operations as possible to improve ergonomics and cut down on archetype moves: ```rust // Old (0.8) commands.spawn() .insert_bundle(SomeBundle::default()) .insert(SomeComponent); // New (0.9) - Option 1 commands.spawn().insert(( SomeBundle::default(), SomeComponent, )) // New (0.9) - Option 2 commands.spawn_bundle(( SomeBundle::default(), SomeComponent, )) ``` ## Next Steps Consider changing `spawn` to accept a bundle and deprecate `spawn_bundle`.
2022-09-21 21:47:53 +00:00
world.entity_mut(e4).insert((A(0), B(0)));
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
assert!(get_filtered::<Added<A>>(&mut world).is_empty());
Reliable change detection (#1471) # Problem Definition The current change tracking (via flags for both components and resources) fails to detect changes made by systems that are scheduled to run earlier in the frame than they are. This issue is discussed at length in [#68](https://github.com/bevyengine/bevy/issues/68) and [#54](https://github.com/bevyengine/bevy/issues/54). This is very much a draft PR, and contributions are welcome and needed. # Criteria 1. Each change is detected at least once, no matter the ordering. 2. Each change is detected at most once, no matter the ordering. 3. Changes should be detected the same frame that they are made. 4. Competitive ergonomics. Ideally does not require opting-in. 5. Low CPU overhead of computation. 6. Memory efficient. This must not increase over time, except where the number of entities / resources does. 7. Changes should not be lost for systems that don't run. 8. A frame needs to act as a pure function. Given the same set of entities / components it needs to produce the same end state without side-effects. **Exact** change-tracking proposals satisfy criteria 1 and 2. **Conservative** change-tracking proposals satisfy criteria 1 but not 2. **Flaky** change tracking proposals satisfy criteria 2 but not 1. # Code Base Navigation There are three types of flags: - `Added`: A piece of data was added to an entity / `Resources`. - `Mutated`: A piece of data was able to be modified, because its `DerefMut` was accessed - `Changed`: The bitwise OR of `Added` and `Changed` The special behavior of `ChangedRes`, with respect to the scheduler is being removed in [#1313](https://github.com/bevyengine/bevy/pull/1313) and does not need to be reproduced. `ChangedRes` and friends can be found in "bevy_ecs/core/resources/resource_query.rs". The `Flags` trait for Components can be found in "bevy_ecs/core/query.rs". `ComponentFlags` are stored in "bevy_ecs/core/archetypes.rs", defined on line 446. # Proposals **Proposal 5 was selected for implementation.** ## Proposal 0: No Change Detection The baseline, where computations are performed on everything regardless of whether it changed. **Type:** Conservative **Pros:** - already implemented - will never miss events - no overhead **Cons:** - tons of repeated work - doesn't allow users to avoid repeating work (or monitoring for other changes) ## Proposal 1: Earlier-This-Tick Change Detection The current approach as of Bevy 0.4. Flags are set, and then flushed at the end of each frame. **Type:** Flaky **Pros:** - already implemented - simple to understand - low memory overhead (2 bits per component) - low time overhead (clear every flag once per frame) **Cons:** - misses systems based on ordering - systems that don't run every frame miss changes - duplicates detection when looping - can lead to unresolvable circular dependencies ## Proposal 2: Two-Tick Change Detection Flags persist for two frames, using a double-buffer system identical to that used in events. A change is observed if it is found in either the current frame's list of changes or the previous frame's. **Type:** Conservative **Pros:** - easy to understand - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - can result in a great deal of duplicated work - systems that don't run every frame miss changes - duplicates detection when looping ## Proposal 3: Last-Tick Change Detection Flags persist for two frames, using a double-buffer system identical to that used in events. A change is observed if it is found in the previous frame's list of changes. **Type:** Exact **Pros:** - exact - easy to understand - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - change detection is always delayed, possibly causing painful chained delays - systems that don't run every frame miss changes - duplicates detection when looping ## Proposal 4: Flag-Doubling Change Detection Combine Proposal 2 and Proposal 3. Differentiate between `JustChanged` (current behavior) and `Changed` (Proposal 3). Pack this data into the flags according to [this implementation proposal](https://github.com/bevyengine/bevy/issues/68#issuecomment-769174804). **Type:** Flaky + Exact **Pros:** - allows users to acc - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - users must specify the type of change detection required - still quite fragile to system ordering effects when using the flaky `JustChanged` form - cannot get immediate + exact results - systems that don't run every frame miss changes - duplicates detection when looping ## [SELECTED] Proposal 5: Generation-Counter Change Detection A global counter is increased after each system is run. Each component saves the time of last mutation, and each system saves the time of last execution. Mutation is detected when the component's counter is greater than the system's counter. Discussed [here](https://github.com/bevyengine/bevy/issues/68#issuecomment-769174804). How to handle addition detection is unsolved; the current proposal is to use the highest bit of the counter as in proposal 1. **Type:** Exact (for mutations), flaky (for additions) **Pros:** - low time overhead (set component counter on access, set system counter after execution) - robust to systems that don't run every frame - robust to systems that loop **Cons:** - moderately complex implementation - must be modified as systems are inserted dynamically - medium memory overhead (4 bytes per component + system) - unsolved addition detection ## Proposal 6: System-Data Change Detection For each system, track which system's changes it has seen. This approach is only worth fully designing and implementing if Proposal 5 fails in some way. **Type:** Exact **Pros:** - exact - conceptually simple **Cons:** - requires storing data on each system - implementation is complex - must be modified as systems are inserted dynamically ## Proposal 7: Total-Order Change Detection Discussed [here](https://github.com/bevyengine/bevy/issues/68#issuecomment-754326523). This proposal is somewhat complicated by the new scheduler, but I believe it should still be conceptually feasible. This approach is only worth fully designing and implementing if Proposal 5 fails in some way. **Type:** Exact **Pros:** - exact - efficient data storage relative to other exact proposals **Cons:** - requires access to the scheduler - complex implementation and difficulty grokking - must be modified as systems are inserted dynamically # Tests - We will need to verify properties 1, 2, 3, 7 and 8. Priority: 1 > 2 = 3 > 8 > 7 - Ideally we can use identical user-facing syntax for all proposals, allowing us to re-use the same syntax for each. - When writing tests, we need to carefully specify order using explicit dependencies. - These tests will need to be duplicated for both components and resources. - We need to be sure to handle cases where ambiguous system orders exist. `changing_system` is always the system that makes the changes, and `detecting_system` always detects the changes. The component / resource changed will be simple boolean wrapper structs. ## Basic Added / Mutated / Changed 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs before `detecting_system` - verify at the end of tick 2 ## At Least Once 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs after `detecting_system` - verify at the end of tick 2 ## At Most Once 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs once before `detecting_system` - increment a counter based on the number of changes detected - verify at the end of tick 2 ## Fast Detection 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs before `detecting_system` - verify at the end of tick 1 ## Ambiguous System Ordering Robustness 2 x 3 x 2 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs [before/after] `detecting_system` in tick 1 - `changing_system` runs [after/before] `detecting_system` in tick 2 ## System Pausing 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs in tick 1, then is disabled by run criteria - `detecting_system` is disabled by run criteria until it is run once during tick 3 - verify at the end of tick 3 ## Addition Causes Mutation 2 design: - Resources vs. Components - `adding_system_1` adds a component / resource - `adding system_2` adds the same component / resource - verify the `Mutated` flag at the end of the tick - verify the `Added` flag at the end of the tick First check tests for: https://github.com/bevyengine/bevy/issues/333 Second check tests for: https://github.com/bevyengine/bevy/issues/1443 ## Changes Made By Commands - `adding_system` runs in Update in tick 1, and sends a command to add a component - `detecting_system` runs in Update in tick 1 and 2, after `adding_system` - We can't detect the changes in tick 1, since they haven't been processed yet - If we were to track these changes as being emitted by `adding_system`, we can't detect the changes in tick 2 either, since `detecting_system` has already run once after `adding_system` :( # Benchmarks See: [general advice](https://github.com/bevyengine/bevy/blob/master/docs/profiling.md), [Criterion crate](https://github.com/bheisler/criterion.rs) There are several critical parameters to vary: 1. entity count (1 to 10^9) 2. fraction of entities that are changed (0% to 100%) 3. cost to perform work on changed entities, i.e. workload (1 ns to 1s) 1 and 2 should be varied between benchmark runs. 3 can be added on computationally. We want to measure: - memory cost - run time We should collect these measurements across several frames (100?) to reduce bootup effects and accurately measure the mean, variance and drift. Entity-component change detection is much more important to benchmark than resource change detection, due to the orders of magnitude higher number of pieces of data. No change detection at all should be included in benchmarks as a second control for cases where missing changes is unacceptable. ## Graphs 1. y: performance, x: log_10(entity count), color: proposal, facet: performance metric. Set cost to perform work to 0. 2. y: run time, x: cost to perform work, color: proposal, facet: fraction changed. Set number of entities to 10^6 3. y: memory, x: frames, color: proposal # Conclusions 1. Is the theoretical categorization of the proposals correct according to our tests? 2. How does the performance of the proposals compare without any load? 3. How does the performance of the proposals compare with realistic loads? 4. At what workload does more exact change tracking become worth the (presumably) higher overhead? 5. When does adding change-detection to save on work become worthwhile? 6. Is there enough divergence in performance between the best solutions in each class to ship more than one change-tracking solution? # Implementation Plan 1. Write a test suite. 2. Verify that tests fail for existing approach. 3. Write a benchmark suite. 4. Get performance numbers for existing approach. 5. Implement, test and benchmark various solutions using a Git branch per proposal. 6. Create a draft PR with all solutions and present results to team. 7. Select a solution and replace existing change detection. Co-authored-by: Brice DAVIER <bricedavier@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2021-03-19 17:53:26 +00:00
assert_eq!(get_filtered::<Changed<A>>(&mut world), vec![e4]);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
assert_eq!(get_filtered::<Added<B>>(&mut world), vec![e4]);
Reliable change detection (#1471) # Problem Definition The current change tracking (via flags for both components and resources) fails to detect changes made by systems that are scheduled to run earlier in the frame than they are. This issue is discussed at length in [#68](https://github.com/bevyengine/bevy/issues/68) and [#54](https://github.com/bevyengine/bevy/issues/54). This is very much a draft PR, and contributions are welcome and needed. # Criteria 1. Each change is detected at least once, no matter the ordering. 2. Each change is detected at most once, no matter the ordering. 3. Changes should be detected the same frame that they are made. 4. Competitive ergonomics. Ideally does not require opting-in. 5. Low CPU overhead of computation. 6. Memory efficient. This must not increase over time, except where the number of entities / resources does. 7. Changes should not be lost for systems that don't run. 8. A frame needs to act as a pure function. Given the same set of entities / components it needs to produce the same end state without side-effects. **Exact** change-tracking proposals satisfy criteria 1 and 2. **Conservative** change-tracking proposals satisfy criteria 1 but not 2. **Flaky** change tracking proposals satisfy criteria 2 but not 1. # Code Base Navigation There are three types of flags: - `Added`: A piece of data was added to an entity / `Resources`. - `Mutated`: A piece of data was able to be modified, because its `DerefMut` was accessed - `Changed`: The bitwise OR of `Added` and `Changed` The special behavior of `ChangedRes`, with respect to the scheduler is being removed in [#1313](https://github.com/bevyengine/bevy/pull/1313) and does not need to be reproduced. `ChangedRes` and friends can be found in "bevy_ecs/core/resources/resource_query.rs". The `Flags` trait for Components can be found in "bevy_ecs/core/query.rs". `ComponentFlags` are stored in "bevy_ecs/core/archetypes.rs", defined on line 446. # Proposals **Proposal 5 was selected for implementation.** ## Proposal 0: No Change Detection The baseline, where computations are performed on everything regardless of whether it changed. **Type:** Conservative **Pros:** - already implemented - will never miss events - no overhead **Cons:** - tons of repeated work - doesn't allow users to avoid repeating work (or monitoring for other changes) ## Proposal 1: Earlier-This-Tick Change Detection The current approach as of Bevy 0.4. Flags are set, and then flushed at the end of each frame. **Type:** Flaky **Pros:** - already implemented - simple to understand - low memory overhead (2 bits per component) - low time overhead (clear every flag once per frame) **Cons:** - misses systems based on ordering - systems that don't run every frame miss changes - duplicates detection when looping - can lead to unresolvable circular dependencies ## Proposal 2: Two-Tick Change Detection Flags persist for two frames, using a double-buffer system identical to that used in events. A change is observed if it is found in either the current frame's list of changes or the previous frame's. **Type:** Conservative **Pros:** - easy to understand - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - can result in a great deal of duplicated work - systems that don't run every frame miss changes - duplicates detection when looping ## Proposal 3: Last-Tick Change Detection Flags persist for two frames, using a double-buffer system identical to that used in events. A change is observed if it is found in the previous frame's list of changes. **Type:** Exact **Pros:** - exact - easy to understand - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - change detection is always delayed, possibly causing painful chained delays - systems that don't run every frame miss changes - duplicates detection when looping ## Proposal 4: Flag-Doubling Change Detection Combine Proposal 2 and Proposal 3. Differentiate between `JustChanged` (current behavior) and `Changed` (Proposal 3). Pack this data into the flags according to [this implementation proposal](https://github.com/bevyengine/bevy/issues/68#issuecomment-769174804). **Type:** Flaky + Exact **Pros:** - allows users to acc - easy to implement - low memory overhead (4 bits per component) - low time overhead (bit mask and shift every flag once per frame) **Cons:** - users must specify the type of change detection required - still quite fragile to system ordering effects when using the flaky `JustChanged` form - cannot get immediate + exact results - systems that don't run every frame miss changes - duplicates detection when looping ## [SELECTED] Proposal 5: Generation-Counter Change Detection A global counter is increased after each system is run. Each component saves the time of last mutation, and each system saves the time of last execution. Mutation is detected when the component's counter is greater than the system's counter. Discussed [here](https://github.com/bevyengine/bevy/issues/68#issuecomment-769174804). How to handle addition detection is unsolved; the current proposal is to use the highest bit of the counter as in proposal 1. **Type:** Exact (for mutations), flaky (for additions) **Pros:** - low time overhead (set component counter on access, set system counter after execution) - robust to systems that don't run every frame - robust to systems that loop **Cons:** - moderately complex implementation - must be modified as systems are inserted dynamically - medium memory overhead (4 bytes per component + system) - unsolved addition detection ## Proposal 6: System-Data Change Detection For each system, track which system's changes it has seen. This approach is only worth fully designing and implementing if Proposal 5 fails in some way. **Type:** Exact **Pros:** - exact - conceptually simple **Cons:** - requires storing data on each system - implementation is complex - must be modified as systems are inserted dynamically ## Proposal 7: Total-Order Change Detection Discussed [here](https://github.com/bevyengine/bevy/issues/68#issuecomment-754326523). This proposal is somewhat complicated by the new scheduler, but I believe it should still be conceptually feasible. This approach is only worth fully designing and implementing if Proposal 5 fails in some way. **Type:** Exact **Pros:** - exact - efficient data storage relative to other exact proposals **Cons:** - requires access to the scheduler - complex implementation and difficulty grokking - must be modified as systems are inserted dynamically # Tests - We will need to verify properties 1, 2, 3, 7 and 8. Priority: 1 > 2 = 3 > 8 > 7 - Ideally we can use identical user-facing syntax for all proposals, allowing us to re-use the same syntax for each. - When writing tests, we need to carefully specify order using explicit dependencies. - These tests will need to be duplicated for both components and resources. - We need to be sure to handle cases where ambiguous system orders exist. `changing_system` is always the system that makes the changes, and `detecting_system` always detects the changes. The component / resource changed will be simple boolean wrapper structs. ## Basic Added / Mutated / Changed 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs before `detecting_system` - verify at the end of tick 2 ## At Least Once 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs after `detecting_system` - verify at the end of tick 2 ## At Most Once 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs once before `detecting_system` - increment a counter based on the number of changes detected - verify at the end of tick 2 ## Fast Detection 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs before `detecting_system` - verify at the end of tick 1 ## Ambiguous System Ordering Robustness 2 x 3 x 2 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs [before/after] `detecting_system` in tick 1 - `changing_system` runs [after/before] `detecting_system` in tick 2 ## System Pausing 2 x 3 design: - Resources vs. Components - Added vs. Changed vs. Mutated - `changing_system` runs in tick 1, then is disabled by run criteria - `detecting_system` is disabled by run criteria until it is run once during tick 3 - verify at the end of tick 3 ## Addition Causes Mutation 2 design: - Resources vs. Components - `adding_system_1` adds a component / resource - `adding system_2` adds the same component / resource - verify the `Mutated` flag at the end of the tick - verify the `Added` flag at the end of the tick First check tests for: https://github.com/bevyengine/bevy/issues/333 Second check tests for: https://github.com/bevyengine/bevy/issues/1443 ## Changes Made By Commands - `adding_system` runs in Update in tick 1, and sends a command to add a component - `detecting_system` runs in Update in tick 1 and 2, after `adding_system` - We can't detect the changes in tick 1, since they haven't been processed yet - If we were to track these changes as being emitted by `adding_system`, we can't detect the changes in tick 2 either, since `detecting_system` has already run once after `adding_system` :( # Benchmarks See: [general advice](https://github.com/bevyengine/bevy/blob/master/docs/profiling.md), [Criterion crate](https://github.com/bheisler/criterion.rs) There are several critical parameters to vary: 1. entity count (1 to 10^9) 2. fraction of entities that are changed (0% to 100%) 3. cost to perform work on changed entities, i.e. workload (1 ns to 1s) 1 and 2 should be varied between benchmark runs. 3 can be added on computationally. We want to measure: - memory cost - run time We should collect these measurements across several frames (100?) to reduce bootup effects and accurately measure the mean, variance and drift. Entity-component change detection is much more important to benchmark than resource change detection, due to the orders of magnitude higher number of pieces of data. No change detection at all should be included in benchmarks as a second control for cases where missing changes is unacceptable. ## Graphs 1. y: performance, x: log_10(entity count), color: proposal, facet: performance metric. Set cost to perform work to 0. 2. y: run time, x: cost to perform work, color: proposal, facet: fraction changed. Set number of entities to 10^6 3. y: memory, x: frames, color: proposal # Conclusions 1. Is the theoretical categorization of the proposals correct according to our tests? 2. How does the performance of the proposals compare without any load? 3. How does the performance of the proposals compare with realistic loads? 4. At what workload does more exact change tracking become worth the (presumably) higher overhead? 5. When does adding change-detection to save on work become worthwhile? 6. Is there enough divergence in performance between the best solutions in each class to ship more than one change-tracking solution? # Implementation Plan 1. Write a test suite. 2. Verify that tests fail for existing approach. 3. Write a benchmark suite. 4. Get performance numbers for existing approach. 5. Implement, test and benchmark various solutions using a Git branch per proposal. 6. Create a draft PR with all solutions and present results to team. 7. Select a solution and replace existing change detection. Co-authored-by: Brice DAVIER <bricedavier@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2021-03-19 17:53:26 +00:00
assert_eq!(get_filtered::<Changed<B>>(&mut world), vec![e4]);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
}
#[test]
fn changed_trackers_sparse() {
let mut world = World::default();
let e1 = world.spawn(SparseStored(0)).id();
let e2 = world.spawn(SparseStored(0)).id();
let e3 = world.spawn(SparseStored(0)).id();
world.spawn(SparseStored(0));
world.clear_trackers();
for (i, mut a) in world
.query::<&mut SparseStored>()
.iter_mut(&mut world)
.enumerate()
{
if i % 2 == 0 {
a.0 += 1;
}
}
Rename `WorldQueryData` & `WorldQueryFilter` to `QueryData` & `QueryFilter` (#10779) # Rename `WorldQueryData` & `WorldQueryFilter` to `QueryData` & `QueryFilter` Fixes #10776 ## Solution Traits `WorldQueryData` & `WorldQueryFilter` were renamed to `QueryData` and `QueryFilter`, respectively. Related Trait types were also renamed. --- ## Changelog - Trait `WorldQueryData` has been renamed to `QueryData`. Derive macro's `QueryData` attribute `world_query_data` has been renamed to `query_data`. - Trait `WorldQueryFilter` has been renamed to `QueryFilter`. Derive macro's `QueryFilter` attribute `world_query_filter` has been renamed to `query_filter`. - Trait's `ExtractComponent` type `Query` has been renamed to `Data`. - Trait's `GetBatchData` types `Query` & `QueryFilter` has been renamed to `Data` & `Filter`, respectively. - Trait's `ExtractInstance` type `Query` has been renamed to `Data`. - Trait's `ViewNode` type `ViewQuery` has been renamed to `ViewData`. - Trait's `RenderCommand` types `ViewWorldQuery` & `ItemWorldQuery` has been renamed to `ViewData` & `ItemData`, respectively. ## Migration Guide Note: if merged before 0.13 is released, this should instead modify the migration guide of #10776 with the updated names. - Rename `WorldQueryData` & `WorldQueryFilter` trait usages to `QueryData` & `QueryFilter` and their respective derive macro attributes `world_query_data` & `world_query_filter` to `query_data` & `query_filter`. - Rename the following trait type usages: - Trait's `ExtractComponent` type `Query` to `Data`. - Trait's `GetBatchData` type `Query` to `Data`. - Trait's `ExtractInstance` type `Query` to `Data`. - Trait's `ViewNode` type `ViewQuery` to `ViewData`' - Trait's `RenderCommand` types `ViewWolrdQuery` & `ItemWorldQuery` to `ViewData` & `ItemData`, respectively. ```rust // Before #[derive(WorldQueryData)] #[world_query_data(derive(Debug))] struct EmptyQuery { empty: (), } // After #[derive(QueryData)] #[query_data(derive(Debug))] struct EmptyQuery { empty: (), } // Before #[derive(WorldQueryFilter)] struct CustomQueryFilter<T: Component, P: Component> { _c: With<ComponentC>, _d: With<ComponentD>, _or: Or<(Added<ComponentC>, Changed<ComponentD>, Without<ComponentZ>)>, _generic_tuple: (With<T>, With<P>), } // After #[derive(QueryFilter)] struct CustomQueryFilter<T: Component, P: Component> { _c: With<ComponentC>, _d: With<ComponentD>, _or: Or<(Added<ComponentC>, Changed<ComponentD>, Without<ComponentZ>)>, _generic_tuple: (With<T>, With<P>), } // Before impl ExtractComponent for ContrastAdaptiveSharpeningSettings { type Query = &'static Self; type Filter = With<Camera>; type Out = (DenoiseCAS, CASUniform); fn extract_component(item: QueryItem<Self::Query>) -> Option<Self::Out> { //... } } // After impl ExtractComponent for ContrastAdaptiveSharpeningSettings { type Data = &'static Self; type Filter = With<Camera>; type Out = (DenoiseCAS, CASUniform); fn extract_component(item: QueryItem<Self::Data>) -> Option<Self::Out> { //... } } // Before impl GetBatchData for MeshPipeline { type Param = SRes<RenderMeshInstances>; type Query = Entity; type QueryFilter = With<Mesh3d>; type CompareData = (MaterialBindGroupId, AssetId<Mesh>); type BufferData = MeshUniform; fn get_batch_data( mesh_instances: &SystemParamItem<Self::Param>, entity: &QueryItem<Self::Query>, ) -> (Self::BufferData, Option<Self::CompareData>) { // .... } } // After impl GetBatchData for MeshPipeline { type Param = SRes<RenderMeshInstances>; type Data = Entity; type Filter = With<Mesh3d>; type CompareData = (MaterialBindGroupId, AssetId<Mesh>); type BufferData = MeshUniform; fn get_batch_data( mesh_instances: &SystemParamItem<Self::Param>, entity: &QueryItem<Self::Data>, ) -> (Self::BufferData, Option<Self::CompareData>) { // .... } } // Before impl<A> ExtractInstance for AssetId<A> where A: Asset, { type Query = Read<Handle<A>>; type Filter = (); fn extract(item: QueryItem<'_, Self::Query>) -> Option<Self> { Some(item.id()) } } // After impl<A> ExtractInstance for AssetId<A> where A: Asset, { type Data = Read<Handle<A>>; type Filter = (); fn extract(item: QueryItem<'_, Self::Data>) -> Option<Self> { Some(item.id()) } } // Before impl ViewNode for PostProcessNode { type ViewQuery = ( &'static ViewTarget, &'static PostProcessSettings, ); fn run( &self, _graph: &mut RenderGraphContext, render_context: &mut RenderContext, (view_target, _post_process_settings): QueryItem<Self::ViewQuery>, world: &World, ) -> Result<(), NodeRunError> { // ... } } // After impl ViewNode for PostProcessNode { type ViewData = ( &'static ViewTarget, &'static PostProcessSettings, ); fn run( &self, _graph: &mut RenderGraphContext, render_context: &mut RenderContext, (view_target, _post_process_settings): QueryItem<Self::ViewData>, world: &World, ) -> Result<(), NodeRunError> { // ... } } // Before impl<P: CachedRenderPipelinePhaseItem> RenderCommand<P> for SetItemPipeline { type Param = SRes<PipelineCache>; type ViewWorldQuery = (); type ItemWorldQuery = (); #[inline] fn render<'w>( item: &P, _view: (), _entity: (), pipeline_cache: SystemParamItem<'w, '_, Self::Param>, pass: &mut TrackedRenderPass<'w>, ) -> RenderCommandResult { // ... } } // After impl<P: CachedRenderPipelinePhaseItem> RenderCommand<P> for SetItemPipeline { type Param = SRes<PipelineCache>; type ViewData = (); type ItemData = (); #[inline] fn render<'w>( item: &P, _view: (), _entity: (), pipeline_cache: SystemParamItem<'w, '_, Self::Param>, pass: &mut TrackedRenderPass<'w>, ) -> RenderCommandResult { // ... } } ```
2023-12-12 19:45:50 +00:00
fn get_filtered<F: QueryFilter>(world: &mut World) -> Vec<Entity> {
world
.query_filtered::<Entity, F>()
.iter(world)
.collect::<Vec<Entity>>()
}
assert_eq!(
get_filtered::<Changed<SparseStored>>(&mut world),
vec![e1, e3]
);
// ensure changing an entity's archetypes also moves its changed state
world.entity_mut(e1).insert(C);
assert_eq!(get_filtered::<Changed<SparseStored>>(&mut world), vec![e3, e1], "changed entities list should not change (although the order will due to archetype moves)");
// spawning a new SparseStored entity should not change existing changed state
world.entity_mut(e1).insert(SparseStored(0));
assert_eq!(
get_filtered::<Changed<SparseStored>>(&mut world),
vec![e3, e1],
"changed entities list should not change"
);
// removing an unchanged entity should not change changed state
assert!(world.despawn(e2));
assert_eq!(
get_filtered::<Changed<SparseStored>>(&mut world),
vec![e3, e1],
"changed entities list should not change"
);
// removing a changed entity should remove it from enumeration
assert!(world.despawn(e1));
assert_eq!(
get_filtered::<Changed<SparseStored>>(&mut world),
vec![e3],
"e1 should no longer be returned"
);
world.clear_trackers();
assert!(get_filtered::<Changed<SparseStored>>(&mut world).is_empty());
let e4 = world.spawn_empty().id();
world.entity_mut(e4).insert(SparseStored(0));
assert_eq!(get_filtered::<Changed<SparseStored>>(&mut world), vec![e4]);
assert_eq!(get_filtered::<Added<SparseStored>>(&mut world), vec![e4]);
world.entity_mut(e4).insert(A(1));
assert_eq!(get_filtered::<Changed<SparseStored>>(&mut world), vec![e4]);
world.clear_trackers();
// ensure inserting multiple components set changed state for all components and set added
// state for non existing components even when changing archetype.
world.entity_mut(e4).insert(SparseStored(0));
assert!(get_filtered::<Added<SparseStored>>(&mut world).is_empty());
assert_eq!(get_filtered::<Changed<SparseStored>>(&mut world), vec![e4]);
}
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
#[test]
fn empty_spawn() {
let mut world = World::default();
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
let e = world.spawn_empty().id();
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
let mut e_mut = world.entity_mut(e);
e_mut.insert(A(0));
assert_eq!(e_mut.get::<A>().unwrap(), &A(0));
}
#[test]
fn reserve_and_spawn() {
let mut world = World::default();
let e = world.entities().reserve_entity();
world.flush();
let mut e_mut = world.entity_mut(e);
e_mut.insert(A(0));
assert_eq!(e_mut.get::<A>().unwrap(), &A(0));
}
#[test]
fn changed_query() {
let mut world = World::default();
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
let e1 = world.spawn((A(0), B(0))).id();
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
fn get_changed(world: &mut World) -> Vec<Entity> {
world
.query_filtered::<Entity, Changed<A>>()
.iter(world)
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
.collect::<Vec<Entity>>()
}
assert_eq!(get_changed(&mut world), vec![e1]);
world.clear_trackers();
assert_eq!(get_changed(&mut world), vec![]);
*world.get_mut(e1).unwrap() = A(1);
assert_eq!(get_changed(&mut world), vec![e1]);
}
#[test]
fn resource() {
Make `Resource` trait opt-in, requiring `#[derive(Resource)]` V2 (#5577) *This PR description is an edited copy of #5007, written by @alice-i-cecile.* # Objective Follow-up to https://github.com/bevyengine/bevy/pull/2254. The `Resource` trait currently has a blanket implementation for all types that meet its bounds. While ergonomic, this results in several drawbacks: * it is possible to make confusing, silent mistakes such as inserting a function pointer (Foo) rather than a value (Foo::Bar) as a resource * it is challenging to discover if a type is intended to be used as a resource * we cannot later add customization options (see the [RFC](https://github.com/bevyengine/rfcs/blob/main/rfcs/27-derive-component.md) for the equivalent choice for Component). * dependencies can use the same Rust type as a resource in invisibly conflicting ways * raw Rust types used as resources cannot preserve privacy appropriately, as anyone able to access that type can read and write to internal values * we cannot capture a definitive list of possible resources to display to users in an editor ## Notes to reviewers * Review this commit-by-commit; there's effectively no back-tracking and there's a lot of churn in some of these commits. *ira: My commits are not as well organized :')* * I've relaxed the bound on Local to Send + Sync + 'static: I don't think these concerns apply there, so this can keep things simple. Storing e.g. a u32 in a Local is fine, because there's a variable name attached explaining what it does. * I think this is a bad place for the Resource trait to live, but I've left it in place to make reviewing easier. IMO that's best tackled with https://github.com/bevyengine/bevy/issues/4981. ## Changelog `Resource` is no longer automatically implemented for all matching types. Instead, use the new `#[derive(Resource)]` macro. ## Migration Guide Add `#[derive(Resource)]` to all types you are using as a resource. If you are using a third party type as a resource, wrap it in a tuple struct to bypass orphan rules. Consider deriving `Deref` and `DerefMut` to improve ergonomics. `ClearColor` no longer implements `Component`. Using `ClearColor` as a component in 0.8 did nothing. Use the `ClearColorConfig` in the `Camera3d` and `Camera2d` components instead. Co-authored-by: Alice <alice.i.cecile@gmail.com> Co-authored-by: Alice Cecile <alice.i.cecile@gmail.com> Co-authored-by: devil-ira <justthecooldude@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-08-08 21:36:35 +00:00
use crate::system::Resource;
#[derive(Resource, PartialEq, Debug)]
struct Num(i32);
#[derive(Resource, PartialEq, Debug)]
struct BigNum(u64);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
let mut world = World::default();
Make `Resource` trait opt-in, requiring `#[derive(Resource)]` V2 (#5577) *This PR description is an edited copy of #5007, written by @alice-i-cecile.* # Objective Follow-up to https://github.com/bevyengine/bevy/pull/2254. The `Resource` trait currently has a blanket implementation for all types that meet its bounds. While ergonomic, this results in several drawbacks: * it is possible to make confusing, silent mistakes such as inserting a function pointer (Foo) rather than a value (Foo::Bar) as a resource * it is challenging to discover if a type is intended to be used as a resource * we cannot later add customization options (see the [RFC](https://github.com/bevyengine/rfcs/blob/main/rfcs/27-derive-component.md) for the equivalent choice for Component). * dependencies can use the same Rust type as a resource in invisibly conflicting ways * raw Rust types used as resources cannot preserve privacy appropriately, as anyone able to access that type can read and write to internal values * we cannot capture a definitive list of possible resources to display to users in an editor ## Notes to reviewers * Review this commit-by-commit; there's effectively no back-tracking and there's a lot of churn in some of these commits. *ira: My commits are not as well organized :')* * I've relaxed the bound on Local to Send + Sync + 'static: I don't think these concerns apply there, so this can keep things simple. Storing e.g. a u32 in a Local is fine, because there's a variable name attached explaining what it does. * I think this is a bad place for the Resource trait to live, but I've left it in place to make reviewing easier. IMO that's best tackled with https://github.com/bevyengine/bevy/issues/4981. ## Changelog `Resource` is no longer automatically implemented for all matching types. Instead, use the new `#[derive(Resource)]` macro. ## Migration Guide Add `#[derive(Resource)]` to all types you are using as a resource. If you are using a third party type as a resource, wrap it in a tuple struct to bypass orphan rules. Consider deriving `Deref` and `DerefMut` to improve ergonomics. `ClearColor` no longer implements `Component`. Using `ClearColor` as a component in 0.8 did nothing. Use the `ClearColorConfig` in the `Camera3d` and `Camera2d` components instead. Co-authored-by: Alice <alice.i.cecile@gmail.com> Co-authored-by: Alice Cecile <alice.i.cecile@gmail.com> Co-authored-by: devil-ira <justthecooldude@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-08-08 21:36:35 +00:00
assert!(world.get_resource::<Num>().is_none());
assert!(!world.contains_resource::<Num>());
assert!(!world.is_resource_added::<Num>());
assert!(!world.is_resource_changed::<Num>());
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
Make `Resource` trait opt-in, requiring `#[derive(Resource)]` V2 (#5577) *This PR description is an edited copy of #5007, written by @alice-i-cecile.* # Objective Follow-up to https://github.com/bevyengine/bevy/pull/2254. The `Resource` trait currently has a blanket implementation for all types that meet its bounds. While ergonomic, this results in several drawbacks: * it is possible to make confusing, silent mistakes such as inserting a function pointer (Foo) rather than a value (Foo::Bar) as a resource * it is challenging to discover if a type is intended to be used as a resource * we cannot later add customization options (see the [RFC](https://github.com/bevyengine/rfcs/blob/main/rfcs/27-derive-component.md) for the equivalent choice for Component). * dependencies can use the same Rust type as a resource in invisibly conflicting ways * raw Rust types used as resources cannot preserve privacy appropriately, as anyone able to access that type can read and write to internal values * we cannot capture a definitive list of possible resources to display to users in an editor ## Notes to reviewers * Review this commit-by-commit; there's effectively no back-tracking and there's a lot of churn in some of these commits. *ira: My commits are not as well organized :')* * I've relaxed the bound on Local to Send + Sync + 'static: I don't think these concerns apply there, so this can keep things simple. Storing e.g. a u32 in a Local is fine, because there's a variable name attached explaining what it does. * I think this is a bad place for the Resource trait to live, but I've left it in place to make reviewing easier. IMO that's best tackled with https://github.com/bevyengine/bevy/issues/4981. ## Changelog `Resource` is no longer automatically implemented for all matching types. Instead, use the new `#[derive(Resource)]` macro. ## Migration Guide Add `#[derive(Resource)]` to all types you are using as a resource. If you are using a third party type as a resource, wrap it in a tuple struct to bypass orphan rules. Consider deriving `Deref` and `DerefMut` to improve ergonomics. `ClearColor` no longer implements `Component`. Using `ClearColor` as a component in 0.8 did nothing. Use the `ClearColorConfig` in the `Camera3d` and `Camera2d` components instead. Co-authored-by: Alice <alice.i.cecile@gmail.com> Co-authored-by: Alice Cecile <alice.i.cecile@gmail.com> Co-authored-by: devil-ira <justthecooldude@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-08-08 21:36:35 +00:00
world.insert_resource(Num(123));
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
let resource_id = world
.components()
Make `Resource` trait opt-in, requiring `#[derive(Resource)]` V2 (#5577) *This PR description is an edited copy of #5007, written by @alice-i-cecile.* # Objective Follow-up to https://github.com/bevyengine/bevy/pull/2254. The `Resource` trait currently has a blanket implementation for all types that meet its bounds. While ergonomic, this results in several drawbacks: * it is possible to make confusing, silent mistakes such as inserting a function pointer (Foo) rather than a value (Foo::Bar) as a resource * it is challenging to discover if a type is intended to be used as a resource * we cannot later add customization options (see the [RFC](https://github.com/bevyengine/rfcs/blob/main/rfcs/27-derive-component.md) for the equivalent choice for Component). * dependencies can use the same Rust type as a resource in invisibly conflicting ways * raw Rust types used as resources cannot preserve privacy appropriately, as anyone able to access that type can read and write to internal values * we cannot capture a definitive list of possible resources to display to users in an editor ## Notes to reviewers * Review this commit-by-commit; there's effectively no back-tracking and there's a lot of churn in some of these commits. *ira: My commits are not as well organized :')* * I've relaxed the bound on Local to Send + Sync + 'static: I don't think these concerns apply there, so this can keep things simple. Storing e.g. a u32 in a Local is fine, because there's a variable name attached explaining what it does. * I think this is a bad place for the Resource trait to live, but I've left it in place to make reviewing easier. IMO that's best tackled with https://github.com/bevyengine/bevy/issues/4981. ## Changelog `Resource` is no longer automatically implemented for all matching types. Instead, use the new `#[derive(Resource)]` macro. ## Migration Guide Add `#[derive(Resource)]` to all types you are using as a resource. If you are using a third party type as a resource, wrap it in a tuple struct to bypass orphan rules. Consider deriving `Deref` and `DerefMut` to improve ergonomics. `ClearColor` no longer implements `Component`. Using `ClearColor` as a component in 0.8 did nothing. Use the `ClearColorConfig` in the `Camera3d` and `Camera2d` components instead. Co-authored-by: Alice <alice.i.cecile@gmail.com> Co-authored-by: Alice Cecile <alice.i.cecile@gmail.com> Co-authored-by: devil-ira <justthecooldude@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-08-08 21:36:35 +00:00
.get_resource_id(TypeId::of::<Num>())
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
.unwrap();
Extract Resources into their own dedicated storage (#4809) # Objective At least partially addresses #6282. Resources are currently stored as a dedicated Resource archetype (ID 1). This allows for easy code reusability, but unnecessarily adds 72 bytes (on 64-bit systems) to the struct that is only used for that one archetype. It also requires several fields to be `pub(crate)` which isn't ideal. This should also remove one sparse-set lookup from fetching, inserting, and removing resources from a `World`. ## Solution - Add `Resources` parallel to `Tables` and `SparseSets` and extract the functionality used by `Archetype` in it. - Remove `unique_components` from `Archetype` - Remove the `pub(crate)` on `Archetype::components`. - Remove `ArchetypeId::RESOURCE` - Remove `Archetypes::resource` and `Archetypes::resource_mut` --- ## Changelog Added: `Resources` type to store resources. Added: `Storages::resource` Removed: `ArchetypeId::RESOURCE` Removed: `Archetypes::resource` and `Archetypes::resources` Removed: `Archetype::unique_components` and `Archetypes::unique_components_mut` ## Migration Guide Resources have been moved to `Resources` under `Storages` in `World`. All code dependent on `Archetype::unique_components(_mut)` should access it via `world.storages().resources()` instead. All APIs accessing the raw data of individual resources (mutable *and* read-only) have been removed as these APIs allowed for unsound unsafe code. All usages of these APIs should be changed to use `World::{get, insert, remove}_resource`.
2022-10-24 13:46:36 +00:00
let archetype_component_id = world.storages().resources.get(resource_id).unwrap().id();
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
Make `Resource` trait opt-in, requiring `#[derive(Resource)]` V2 (#5577) *This PR description is an edited copy of #5007, written by @alice-i-cecile.* # Objective Follow-up to https://github.com/bevyengine/bevy/pull/2254. The `Resource` trait currently has a blanket implementation for all types that meet its bounds. While ergonomic, this results in several drawbacks: * it is possible to make confusing, silent mistakes such as inserting a function pointer (Foo) rather than a value (Foo::Bar) as a resource * it is challenging to discover if a type is intended to be used as a resource * we cannot later add customization options (see the [RFC](https://github.com/bevyengine/rfcs/blob/main/rfcs/27-derive-component.md) for the equivalent choice for Component). * dependencies can use the same Rust type as a resource in invisibly conflicting ways * raw Rust types used as resources cannot preserve privacy appropriately, as anyone able to access that type can read and write to internal values * we cannot capture a definitive list of possible resources to display to users in an editor ## Notes to reviewers * Review this commit-by-commit; there's effectively no back-tracking and there's a lot of churn in some of these commits. *ira: My commits are not as well organized :')* * I've relaxed the bound on Local to Send + Sync + 'static: I don't think these concerns apply there, so this can keep things simple. Storing e.g. a u32 in a Local is fine, because there's a variable name attached explaining what it does. * I think this is a bad place for the Resource trait to live, but I've left it in place to make reviewing easier. IMO that's best tackled with https://github.com/bevyengine/bevy/issues/4981. ## Changelog `Resource` is no longer automatically implemented for all matching types. Instead, use the new `#[derive(Resource)]` macro. ## Migration Guide Add `#[derive(Resource)]` to all types you are using as a resource. If you are using a third party type as a resource, wrap it in a tuple struct to bypass orphan rules. Consider deriving `Deref` and `DerefMut` to improve ergonomics. `ClearColor` no longer implements `Component`. Using `ClearColor` as a component in 0.8 did nothing. Use the `ClearColorConfig` in the `Camera3d` and `Camera2d` components instead. Co-authored-by: Alice <alice.i.cecile@gmail.com> Co-authored-by: Alice Cecile <alice.i.cecile@gmail.com> Co-authored-by: devil-ira <justthecooldude@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-08-08 21:36:35 +00:00
assert_eq!(world.resource::<Num>().0, 123);
assert!(world.contains_resource::<Num>());
assert!(world.is_resource_added::<Num>());
assert!(world.is_resource_changed::<Num>());
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
Make `Resource` trait opt-in, requiring `#[derive(Resource)]` V2 (#5577) *This PR description is an edited copy of #5007, written by @alice-i-cecile.* # Objective Follow-up to https://github.com/bevyengine/bevy/pull/2254. The `Resource` trait currently has a blanket implementation for all types that meet its bounds. While ergonomic, this results in several drawbacks: * it is possible to make confusing, silent mistakes such as inserting a function pointer (Foo) rather than a value (Foo::Bar) as a resource * it is challenging to discover if a type is intended to be used as a resource * we cannot later add customization options (see the [RFC](https://github.com/bevyengine/rfcs/blob/main/rfcs/27-derive-component.md) for the equivalent choice for Component). * dependencies can use the same Rust type as a resource in invisibly conflicting ways * raw Rust types used as resources cannot preserve privacy appropriately, as anyone able to access that type can read and write to internal values * we cannot capture a definitive list of possible resources to display to users in an editor ## Notes to reviewers * Review this commit-by-commit; there's effectively no back-tracking and there's a lot of churn in some of these commits. *ira: My commits are not as well organized :')* * I've relaxed the bound on Local to Send + Sync + 'static: I don't think these concerns apply there, so this can keep things simple. Storing e.g. a u32 in a Local is fine, because there's a variable name attached explaining what it does. * I think this is a bad place for the Resource trait to live, but I've left it in place to make reviewing easier. IMO that's best tackled with https://github.com/bevyengine/bevy/issues/4981. ## Changelog `Resource` is no longer automatically implemented for all matching types. Instead, use the new `#[derive(Resource)]` macro. ## Migration Guide Add `#[derive(Resource)]` to all types you are using as a resource. If you are using a third party type as a resource, wrap it in a tuple struct to bypass orphan rules. Consider deriving `Deref` and `DerefMut` to improve ergonomics. `ClearColor` no longer implements `Component`. Using `ClearColor` as a component in 0.8 did nothing. Use the `ClearColorConfig` in the `Camera3d` and `Camera2d` components instead. Co-authored-by: Alice <alice.i.cecile@gmail.com> Co-authored-by: Alice Cecile <alice.i.cecile@gmail.com> Co-authored-by: devil-ira <justthecooldude@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-08-08 21:36:35 +00:00
world.insert_resource(BigNum(456));
assert_eq!(world.resource::<BigNum>().0, 456u64);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
Make `Resource` trait opt-in, requiring `#[derive(Resource)]` V2 (#5577) *This PR description is an edited copy of #5007, written by @alice-i-cecile.* # Objective Follow-up to https://github.com/bevyengine/bevy/pull/2254. The `Resource` trait currently has a blanket implementation for all types that meet its bounds. While ergonomic, this results in several drawbacks: * it is possible to make confusing, silent mistakes such as inserting a function pointer (Foo) rather than a value (Foo::Bar) as a resource * it is challenging to discover if a type is intended to be used as a resource * we cannot later add customization options (see the [RFC](https://github.com/bevyengine/rfcs/blob/main/rfcs/27-derive-component.md) for the equivalent choice for Component). * dependencies can use the same Rust type as a resource in invisibly conflicting ways * raw Rust types used as resources cannot preserve privacy appropriately, as anyone able to access that type can read and write to internal values * we cannot capture a definitive list of possible resources to display to users in an editor ## Notes to reviewers * Review this commit-by-commit; there's effectively no back-tracking and there's a lot of churn in some of these commits. *ira: My commits are not as well organized :')* * I've relaxed the bound on Local to Send + Sync + 'static: I don't think these concerns apply there, so this can keep things simple. Storing e.g. a u32 in a Local is fine, because there's a variable name attached explaining what it does. * I think this is a bad place for the Resource trait to live, but I've left it in place to make reviewing easier. IMO that's best tackled with https://github.com/bevyengine/bevy/issues/4981. ## Changelog `Resource` is no longer automatically implemented for all matching types. Instead, use the new `#[derive(Resource)]` macro. ## Migration Guide Add `#[derive(Resource)]` to all types you are using as a resource. If you are using a third party type as a resource, wrap it in a tuple struct to bypass orphan rules. Consider deriving `Deref` and `DerefMut` to improve ergonomics. `ClearColor` no longer implements `Component`. Using `ClearColor` as a component in 0.8 did nothing. Use the `ClearColorConfig` in the `Camera3d` and `Camera2d` components instead. Co-authored-by: Alice <alice.i.cecile@gmail.com> Co-authored-by: Alice Cecile <alice.i.cecile@gmail.com> Co-authored-by: devil-ira <justthecooldude@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-08-08 21:36:35 +00:00
world.insert_resource(BigNum(789));
assert_eq!(world.resource::<BigNum>().0, 789);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
{
Make `Resource` trait opt-in, requiring `#[derive(Resource)]` V2 (#5577) *This PR description is an edited copy of #5007, written by @alice-i-cecile.* # Objective Follow-up to https://github.com/bevyengine/bevy/pull/2254. The `Resource` trait currently has a blanket implementation for all types that meet its bounds. While ergonomic, this results in several drawbacks: * it is possible to make confusing, silent mistakes such as inserting a function pointer (Foo) rather than a value (Foo::Bar) as a resource * it is challenging to discover if a type is intended to be used as a resource * we cannot later add customization options (see the [RFC](https://github.com/bevyengine/rfcs/blob/main/rfcs/27-derive-component.md) for the equivalent choice for Component). * dependencies can use the same Rust type as a resource in invisibly conflicting ways * raw Rust types used as resources cannot preserve privacy appropriately, as anyone able to access that type can read and write to internal values * we cannot capture a definitive list of possible resources to display to users in an editor ## Notes to reviewers * Review this commit-by-commit; there's effectively no back-tracking and there's a lot of churn in some of these commits. *ira: My commits are not as well organized :')* * I've relaxed the bound on Local to Send + Sync + 'static: I don't think these concerns apply there, so this can keep things simple. Storing e.g. a u32 in a Local is fine, because there's a variable name attached explaining what it does. * I think this is a bad place for the Resource trait to live, but I've left it in place to make reviewing easier. IMO that's best tackled with https://github.com/bevyengine/bevy/issues/4981. ## Changelog `Resource` is no longer automatically implemented for all matching types. Instead, use the new `#[derive(Resource)]` macro. ## Migration Guide Add `#[derive(Resource)]` to all types you are using as a resource. If you are using a third party type as a resource, wrap it in a tuple struct to bypass orphan rules. Consider deriving `Deref` and `DerefMut` to improve ergonomics. `ClearColor` no longer implements `Component`. Using `ClearColor` as a component in 0.8 did nothing. Use the `ClearColorConfig` in the `Camera3d` and `Camera2d` components instead. Co-authored-by: Alice <alice.i.cecile@gmail.com> Co-authored-by: Alice Cecile <alice.i.cecile@gmail.com> Co-authored-by: devil-ira <justthecooldude@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-08-08 21:36:35 +00:00
let mut value = world.resource_mut::<BigNum>();
assert_eq!(value.0, 789);
value.0 = 10;
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
}
assert_eq!(
Make `Resource` trait opt-in, requiring `#[derive(Resource)]` V2 (#5577) *This PR description is an edited copy of #5007, written by @alice-i-cecile.* # Objective Follow-up to https://github.com/bevyengine/bevy/pull/2254. The `Resource` trait currently has a blanket implementation for all types that meet its bounds. While ergonomic, this results in several drawbacks: * it is possible to make confusing, silent mistakes such as inserting a function pointer (Foo) rather than a value (Foo::Bar) as a resource * it is challenging to discover if a type is intended to be used as a resource * we cannot later add customization options (see the [RFC](https://github.com/bevyengine/rfcs/blob/main/rfcs/27-derive-component.md) for the equivalent choice for Component). * dependencies can use the same Rust type as a resource in invisibly conflicting ways * raw Rust types used as resources cannot preserve privacy appropriately, as anyone able to access that type can read and write to internal values * we cannot capture a definitive list of possible resources to display to users in an editor ## Notes to reviewers * Review this commit-by-commit; there's effectively no back-tracking and there's a lot of churn in some of these commits. *ira: My commits are not as well organized :')* * I've relaxed the bound on Local to Send + Sync + 'static: I don't think these concerns apply there, so this can keep things simple. Storing e.g. a u32 in a Local is fine, because there's a variable name attached explaining what it does. * I think this is a bad place for the Resource trait to live, but I've left it in place to make reviewing easier. IMO that's best tackled with https://github.com/bevyengine/bevy/issues/4981. ## Changelog `Resource` is no longer automatically implemented for all matching types. Instead, use the new `#[derive(Resource)]` macro. ## Migration Guide Add `#[derive(Resource)]` to all types you are using as a resource. If you are using a third party type as a resource, wrap it in a tuple struct to bypass orphan rules. Consider deriving `Deref` and `DerefMut` to improve ergonomics. `ClearColor` no longer implements `Component`. Using `ClearColor` as a component in 0.8 did nothing. Use the `ClearColorConfig` in the `Camera3d` and `Camera2d` components instead. Co-authored-by: Alice <alice.i.cecile@gmail.com> Co-authored-by: Alice Cecile <alice.i.cecile@gmail.com> Co-authored-by: devil-ira <justthecooldude@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-08-08 21:36:35 +00:00
world.resource::<BigNum>().0,
10,
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
"resource changes are preserved"
);
assert_eq!(
Make `Resource` trait opt-in, requiring `#[derive(Resource)]` V2 (#5577) *This PR description is an edited copy of #5007, written by @alice-i-cecile.* # Objective Follow-up to https://github.com/bevyengine/bevy/pull/2254. The `Resource` trait currently has a blanket implementation for all types that meet its bounds. While ergonomic, this results in several drawbacks: * it is possible to make confusing, silent mistakes such as inserting a function pointer (Foo) rather than a value (Foo::Bar) as a resource * it is challenging to discover if a type is intended to be used as a resource * we cannot later add customization options (see the [RFC](https://github.com/bevyengine/rfcs/blob/main/rfcs/27-derive-component.md) for the equivalent choice for Component). * dependencies can use the same Rust type as a resource in invisibly conflicting ways * raw Rust types used as resources cannot preserve privacy appropriately, as anyone able to access that type can read and write to internal values * we cannot capture a definitive list of possible resources to display to users in an editor ## Notes to reviewers * Review this commit-by-commit; there's effectively no back-tracking and there's a lot of churn in some of these commits. *ira: My commits are not as well organized :')* * I've relaxed the bound on Local to Send + Sync + 'static: I don't think these concerns apply there, so this can keep things simple. Storing e.g. a u32 in a Local is fine, because there's a variable name attached explaining what it does. * I think this is a bad place for the Resource trait to live, but I've left it in place to make reviewing easier. IMO that's best tackled with https://github.com/bevyengine/bevy/issues/4981. ## Changelog `Resource` is no longer automatically implemented for all matching types. Instead, use the new `#[derive(Resource)]` macro. ## Migration Guide Add `#[derive(Resource)]` to all types you are using as a resource. If you are using a third party type as a resource, wrap it in a tuple struct to bypass orphan rules. Consider deriving `Deref` and `DerefMut` to improve ergonomics. `ClearColor` no longer implements `Component`. Using `ClearColor` as a component in 0.8 did nothing. Use the `ClearColorConfig` in the `Camera3d` and `Camera2d` components instead. Co-authored-by: Alice <alice.i.cecile@gmail.com> Co-authored-by: Alice Cecile <alice.i.cecile@gmail.com> Co-authored-by: devil-ira <justthecooldude@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-08-08 21:36:35 +00:00
world.remove_resource::<BigNum>(),
Some(BigNum(10)),
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
"removed resource has the correct value"
);
assert_eq!(
Make `Resource` trait opt-in, requiring `#[derive(Resource)]` V2 (#5577) *This PR description is an edited copy of #5007, written by @alice-i-cecile.* # Objective Follow-up to https://github.com/bevyengine/bevy/pull/2254. The `Resource` trait currently has a blanket implementation for all types that meet its bounds. While ergonomic, this results in several drawbacks: * it is possible to make confusing, silent mistakes such as inserting a function pointer (Foo) rather than a value (Foo::Bar) as a resource * it is challenging to discover if a type is intended to be used as a resource * we cannot later add customization options (see the [RFC](https://github.com/bevyengine/rfcs/blob/main/rfcs/27-derive-component.md) for the equivalent choice for Component). * dependencies can use the same Rust type as a resource in invisibly conflicting ways * raw Rust types used as resources cannot preserve privacy appropriately, as anyone able to access that type can read and write to internal values * we cannot capture a definitive list of possible resources to display to users in an editor ## Notes to reviewers * Review this commit-by-commit; there's effectively no back-tracking and there's a lot of churn in some of these commits. *ira: My commits are not as well organized :')* * I've relaxed the bound on Local to Send + Sync + 'static: I don't think these concerns apply there, so this can keep things simple. Storing e.g. a u32 in a Local is fine, because there's a variable name attached explaining what it does. * I think this is a bad place for the Resource trait to live, but I've left it in place to make reviewing easier. IMO that's best tackled with https://github.com/bevyengine/bevy/issues/4981. ## Changelog `Resource` is no longer automatically implemented for all matching types. Instead, use the new `#[derive(Resource)]` macro. ## Migration Guide Add `#[derive(Resource)]` to all types you are using as a resource. If you are using a third party type as a resource, wrap it in a tuple struct to bypass orphan rules. Consider deriving `Deref` and `DerefMut` to improve ergonomics. `ClearColor` no longer implements `Component`. Using `ClearColor` as a component in 0.8 did nothing. Use the `ClearColorConfig` in the `Camera3d` and `Camera2d` components instead. Co-authored-by: Alice <alice.i.cecile@gmail.com> Co-authored-by: Alice Cecile <alice.i.cecile@gmail.com> Co-authored-by: devil-ira <justthecooldude@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-08-08 21:36:35 +00:00
world.get_resource::<BigNum>(),
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
None,
"removed resource no longer exists"
);
assert_eq!(
Make `Resource` trait opt-in, requiring `#[derive(Resource)]` V2 (#5577) *This PR description is an edited copy of #5007, written by @alice-i-cecile.* # Objective Follow-up to https://github.com/bevyengine/bevy/pull/2254. The `Resource` trait currently has a blanket implementation for all types that meet its bounds. While ergonomic, this results in several drawbacks: * it is possible to make confusing, silent mistakes such as inserting a function pointer (Foo) rather than a value (Foo::Bar) as a resource * it is challenging to discover if a type is intended to be used as a resource * we cannot later add customization options (see the [RFC](https://github.com/bevyengine/rfcs/blob/main/rfcs/27-derive-component.md) for the equivalent choice for Component). * dependencies can use the same Rust type as a resource in invisibly conflicting ways * raw Rust types used as resources cannot preserve privacy appropriately, as anyone able to access that type can read and write to internal values * we cannot capture a definitive list of possible resources to display to users in an editor ## Notes to reviewers * Review this commit-by-commit; there's effectively no back-tracking and there's a lot of churn in some of these commits. *ira: My commits are not as well organized :')* * I've relaxed the bound on Local to Send + Sync + 'static: I don't think these concerns apply there, so this can keep things simple. Storing e.g. a u32 in a Local is fine, because there's a variable name attached explaining what it does. * I think this is a bad place for the Resource trait to live, but I've left it in place to make reviewing easier. IMO that's best tackled with https://github.com/bevyengine/bevy/issues/4981. ## Changelog `Resource` is no longer automatically implemented for all matching types. Instead, use the new `#[derive(Resource)]` macro. ## Migration Guide Add `#[derive(Resource)]` to all types you are using as a resource. If you are using a third party type as a resource, wrap it in a tuple struct to bypass orphan rules. Consider deriving `Deref` and `DerefMut` to improve ergonomics. `ClearColor` no longer implements `Component`. Using `ClearColor` as a component in 0.8 did nothing. Use the `ClearColorConfig` in the `Camera3d` and `Camera2d` components instead. Co-authored-by: Alice <alice.i.cecile@gmail.com> Co-authored-by: Alice Cecile <alice.i.cecile@gmail.com> Co-authored-by: devil-ira <justthecooldude@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-08-08 21:36:35 +00:00
world.remove_resource::<BigNum>(),
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
None,
"double remove returns nothing"
);
Make `Resource` trait opt-in, requiring `#[derive(Resource)]` V2 (#5577) *This PR description is an edited copy of #5007, written by @alice-i-cecile.* # Objective Follow-up to https://github.com/bevyengine/bevy/pull/2254. The `Resource` trait currently has a blanket implementation for all types that meet its bounds. While ergonomic, this results in several drawbacks: * it is possible to make confusing, silent mistakes such as inserting a function pointer (Foo) rather than a value (Foo::Bar) as a resource * it is challenging to discover if a type is intended to be used as a resource * we cannot later add customization options (see the [RFC](https://github.com/bevyengine/rfcs/blob/main/rfcs/27-derive-component.md) for the equivalent choice for Component). * dependencies can use the same Rust type as a resource in invisibly conflicting ways * raw Rust types used as resources cannot preserve privacy appropriately, as anyone able to access that type can read and write to internal values * we cannot capture a definitive list of possible resources to display to users in an editor ## Notes to reviewers * Review this commit-by-commit; there's effectively no back-tracking and there's a lot of churn in some of these commits. *ira: My commits are not as well organized :')* * I've relaxed the bound on Local to Send + Sync + 'static: I don't think these concerns apply there, so this can keep things simple. Storing e.g. a u32 in a Local is fine, because there's a variable name attached explaining what it does. * I think this is a bad place for the Resource trait to live, but I've left it in place to make reviewing easier. IMO that's best tackled with https://github.com/bevyengine/bevy/issues/4981. ## Changelog `Resource` is no longer automatically implemented for all matching types. Instead, use the new `#[derive(Resource)]` macro. ## Migration Guide Add `#[derive(Resource)]` to all types you are using as a resource. If you are using a third party type as a resource, wrap it in a tuple struct to bypass orphan rules. Consider deriving `Deref` and `DerefMut` to improve ergonomics. `ClearColor` no longer implements `Component`. Using `ClearColor` as a component in 0.8 did nothing. Use the `ClearColorConfig` in the `Camera3d` and `Camera2d` components instead. Co-authored-by: Alice <alice.i.cecile@gmail.com> Co-authored-by: Alice Cecile <alice.i.cecile@gmail.com> Co-authored-by: devil-ira <justthecooldude@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-08-08 21:36:35 +00:00
world.insert_resource(BigNum(1));
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
assert_eq!(
Make `Resource` trait opt-in, requiring `#[derive(Resource)]` V2 (#5577) *This PR description is an edited copy of #5007, written by @alice-i-cecile.* # Objective Follow-up to https://github.com/bevyengine/bevy/pull/2254. The `Resource` trait currently has a blanket implementation for all types that meet its bounds. While ergonomic, this results in several drawbacks: * it is possible to make confusing, silent mistakes such as inserting a function pointer (Foo) rather than a value (Foo::Bar) as a resource * it is challenging to discover if a type is intended to be used as a resource * we cannot later add customization options (see the [RFC](https://github.com/bevyengine/rfcs/blob/main/rfcs/27-derive-component.md) for the equivalent choice for Component). * dependencies can use the same Rust type as a resource in invisibly conflicting ways * raw Rust types used as resources cannot preserve privacy appropriately, as anyone able to access that type can read and write to internal values * we cannot capture a definitive list of possible resources to display to users in an editor ## Notes to reviewers * Review this commit-by-commit; there's effectively no back-tracking and there's a lot of churn in some of these commits. *ira: My commits are not as well organized :')* * I've relaxed the bound on Local to Send + Sync + 'static: I don't think these concerns apply there, so this can keep things simple. Storing e.g. a u32 in a Local is fine, because there's a variable name attached explaining what it does. * I think this is a bad place for the Resource trait to live, but I've left it in place to make reviewing easier. IMO that's best tackled with https://github.com/bevyengine/bevy/issues/4981. ## Changelog `Resource` is no longer automatically implemented for all matching types. Instead, use the new `#[derive(Resource)]` macro. ## Migration Guide Add `#[derive(Resource)]` to all types you are using as a resource. If you are using a third party type as a resource, wrap it in a tuple struct to bypass orphan rules. Consider deriving `Deref` and `DerefMut` to improve ergonomics. `ClearColor` no longer implements `Component`. Using `ClearColor` as a component in 0.8 did nothing. Use the `ClearColorConfig` in the `Camera3d` and `Camera2d` components instead. Co-authored-by: Alice <alice.i.cecile@gmail.com> Co-authored-by: Alice Cecile <alice.i.cecile@gmail.com> Co-authored-by: devil-ira <justthecooldude@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-08-08 21:36:35 +00:00
world.get_resource::<BigNum>(),
Some(&BigNum(1)),
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
"re-inserting resources works"
);
assert_eq!(
Make `Resource` trait opt-in, requiring `#[derive(Resource)]` V2 (#5577) *This PR description is an edited copy of #5007, written by @alice-i-cecile.* # Objective Follow-up to https://github.com/bevyengine/bevy/pull/2254. The `Resource` trait currently has a blanket implementation for all types that meet its bounds. While ergonomic, this results in several drawbacks: * it is possible to make confusing, silent mistakes such as inserting a function pointer (Foo) rather than a value (Foo::Bar) as a resource * it is challenging to discover if a type is intended to be used as a resource * we cannot later add customization options (see the [RFC](https://github.com/bevyengine/rfcs/blob/main/rfcs/27-derive-component.md) for the equivalent choice for Component). * dependencies can use the same Rust type as a resource in invisibly conflicting ways * raw Rust types used as resources cannot preserve privacy appropriately, as anyone able to access that type can read and write to internal values * we cannot capture a definitive list of possible resources to display to users in an editor ## Notes to reviewers * Review this commit-by-commit; there's effectively no back-tracking and there's a lot of churn in some of these commits. *ira: My commits are not as well organized :')* * I've relaxed the bound on Local to Send + Sync + 'static: I don't think these concerns apply there, so this can keep things simple. Storing e.g. a u32 in a Local is fine, because there's a variable name attached explaining what it does. * I think this is a bad place for the Resource trait to live, but I've left it in place to make reviewing easier. IMO that's best tackled with https://github.com/bevyengine/bevy/issues/4981. ## Changelog `Resource` is no longer automatically implemented for all matching types. Instead, use the new `#[derive(Resource)]` macro. ## Migration Guide Add `#[derive(Resource)]` to all types you are using as a resource. If you are using a third party type as a resource, wrap it in a tuple struct to bypass orphan rules. Consider deriving `Deref` and `DerefMut` to improve ergonomics. `ClearColor` no longer implements `Component`. Using `ClearColor` as a component in 0.8 did nothing. Use the `ClearColorConfig` in the `Camera3d` and `Camera2d` components instead. Co-authored-by: Alice <alice.i.cecile@gmail.com> Co-authored-by: Alice Cecile <alice.i.cecile@gmail.com> Co-authored-by: devil-ira <justthecooldude@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-08-08 21:36:35 +00:00
world.get_resource::<Num>(),
Some(&Num(123)),
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
"other resources are unaffected"
);
let current_resource_id = world
.components()
Make `Resource` trait opt-in, requiring `#[derive(Resource)]` V2 (#5577) *This PR description is an edited copy of #5007, written by @alice-i-cecile.* # Objective Follow-up to https://github.com/bevyengine/bevy/pull/2254. The `Resource` trait currently has a blanket implementation for all types that meet its bounds. While ergonomic, this results in several drawbacks: * it is possible to make confusing, silent mistakes such as inserting a function pointer (Foo) rather than a value (Foo::Bar) as a resource * it is challenging to discover if a type is intended to be used as a resource * we cannot later add customization options (see the [RFC](https://github.com/bevyengine/rfcs/blob/main/rfcs/27-derive-component.md) for the equivalent choice for Component). * dependencies can use the same Rust type as a resource in invisibly conflicting ways * raw Rust types used as resources cannot preserve privacy appropriately, as anyone able to access that type can read and write to internal values * we cannot capture a definitive list of possible resources to display to users in an editor ## Notes to reviewers * Review this commit-by-commit; there's effectively no back-tracking and there's a lot of churn in some of these commits. *ira: My commits are not as well organized :')* * I've relaxed the bound on Local to Send + Sync + 'static: I don't think these concerns apply there, so this can keep things simple. Storing e.g. a u32 in a Local is fine, because there's a variable name attached explaining what it does. * I think this is a bad place for the Resource trait to live, but I've left it in place to make reviewing easier. IMO that's best tackled with https://github.com/bevyengine/bevy/issues/4981. ## Changelog `Resource` is no longer automatically implemented for all matching types. Instead, use the new `#[derive(Resource)]` macro. ## Migration Guide Add `#[derive(Resource)]` to all types you are using as a resource. If you are using a third party type as a resource, wrap it in a tuple struct to bypass orphan rules. Consider deriving `Deref` and `DerefMut` to improve ergonomics. `ClearColor` no longer implements `Component`. Using `ClearColor` as a component in 0.8 did nothing. Use the `ClearColorConfig` in the `Camera3d` and `Camera2d` components instead. Co-authored-by: Alice <alice.i.cecile@gmail.com> Co-authored-by: Alice Cecile <alice.i.cecile@gmail.com> Co-authored-by: devil-ira <justthecooldude@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-08-08 21:36:35 +00:00
.get_resource_id(TypeId::of::<Num>())
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
.unwrap();
assert_eq!(
resource_id, current_resource_id,
"resource id does not change after removing / re-adding"
);
Extract Resources into their own dedicated storage (#4809) # Objective At least partially addresses #6282. Resources are currently stored as a dedicated Resource archetype (ID 1). This allows for easy code reusability, but unnecessarily adds 72 bytes (on 64-bit systems) to the struct that is only used for that one archetype. It also requires several fields to be `pub(crate)` which isn't ideal. This should also remove one sparse-set lookup from fetching, inserting, and removing resources from a `World`. ## Solution - Add `Resources` parallel to `Tables` and `SparseSets` and extract the functionality used by `Archetype` in it. - Remove `unique_components` from `Archetype` - Remove the `pub(crate)` on `Archetype::components`. - Remove `ArchetypeId::RESOURCE` - Remove `Archetypes::resource` and `Archetypes::resource_mut` --- ## Changelog Added: `Resources` type to store resources. Added: `Storages::resource` Removed: `ArchetypeId::RESOURCE` Removed: `Archetypes::resource` and `Archetypes::resources` Removed: `Archetype::unique_components` and `Archetypes::unique_components_mut` ## Migration Guide Resources have been moved to `Resources` under `Storages` in `World`. All code dependent on `Archetype::unique_components(_mut)` should access it via `world.storages().resources()` instead. All APIs accessing the raw data of individual resources (mutable *and* read-only) have been removed as these APIs allowed for unsound unsafe code. All usages of these APIs should be changed to use `World::{get, insert, remove}_resource`.
2022-10-24 13:46:36 +00:00
let current_archetype_component_id =
world.storages().resources.get(resource_id).unwrap().id();
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
assert_eq!(
archetype_component_id, current_archetype_component_id,
"resource archetype component id does not change after removing / re-adding"
);
}
#[test]
fn remove() {
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
let mut world = World::default();
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
let e1 = world.spawn((A(1), B(1), TableStored("a"))).id();
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
let mut e = world.entity_mut(e1);
assert_eq!(e.get::<TableStored>(), Some(&TableStored("a")));
assert_eq!(e.get::<A>(), Some(&A(1)));
assert_eq!(e.get::<B>(), Some(&B(1)));
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
assert_eq!(
e.get::<C>(),
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
None,
"C is not in the entity, so it should not exist"
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
);
e.remove::<(A, B, C)>();
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
assert_eq!(
e.get::<TableStored>(),
Some(&TableStored("a")),
"TableStored is not in the removed bundle, so it should exist"
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
);
assert_eq!(
e.get::<A>(),
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
None,
Make `Resource` trait opt-in, requiring `#[derive(Resource)]` V2 (#5577) *This PR description is an edited copy of #5007, written by @alice-i-cecile.* # Objective Follow-up to https://github.com/bevyengine/bevy/pull/2254. The `Resource` trait currently has a blanket implementation for all types that meet its bounds. While ergonomic, this results in several drawbacks: * it is possible to make confusing, silent mistakes such as inserting a function pointer (Foo) rather than a value (Foo::Bar) as a resource * it is challenging to discover if a type is intended to be used as a resource * we cannot later add customization options (see the [RFC](https://github.com/bevyengine/rfcs/blob/main/rfcs/27-derive-component.md) for the equivalent choice for Component). * dependencies can use the same Rust type as a resource in invisibly conflicting ways * raw Rust types used as resources cannot preserve privacy appropriately, as anyone able to access that type can read and write to internal values * we cannot capture a definitive list of possible resources to display to users in an editor ## Notes to reviewers * Review this commit-by-commit; there's effectively no back-tracking and there's a lot of churn in some of these commits. *ira: My commits are not as well organized :')* * I've relaxed the bound on Local to Send + Sync + 'static: I don't think these concerns apply there, so this can keep things simple. Storing e.g. a u32 in a Local is fine, because there's a variable name attached explaining what it does. * I think this is a bad place for the Resource trait to live, but I've left it in place to make reviewing easier. IMO that's best tackled with https://github.com/bevyengine/bevy/issues/4981. ## Changelog `Resource` is no longer automatically implemented for all matching types. Instead, use the new `#[derive(Resource)]` macro. ## Migration Guide Add `#[derive(Resource)]` to all types you are using as a resource. If you are using a third party type as a resource, wrap it in a tuple struct to bypass orphan rules. Consider deriving `Deref` and `DerefMut` to improve ergonomics. `ClearColor` no longer implements `Component`. Using `ClearColor` as a component in 0.8 did nothing. Use the `ClearColorConfig` in the `Camera3d` and `Camera2d` components instead. Co-authored-by: Alice <alice.i.cecile@gmail.com> Co-authored-by: Alice Cecile <alice.i.cecile@gmail.com> Co-authored-by: devil-ira <justthecooldude@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-08-08 21:36:35 +00:00
"Num is in the removed bundle, so it should not exist"
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
);
assert_eq!(
e.get::<B>(),
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
None,
"f64 is in the removed bundle, so it should not exist"
);
assert_eq!(
e.get::<C>(),
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
None,
"usize is in the removed bundle, so it should not exist"
);
}
#[test]
fn take() {
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
let mut world = World::default();
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
world.spawn((A(1), B(1), TableStored("1")));
let e2 = world.spawn((A(2), B(2), TableStored("2"))).id();
world.spawn((A(3), B(3), TableStored("3")));
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
let mut query = world.query::<(&B, &TableStored)>();
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
let results = query
.iter(&world)
.map(|(a, b)| (a.0, b.0))
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
.collect::<Vec<_>>();
assert_eq!(results, vec![(1, "1"), (2, "2"), (3, "3"),]);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
let removed_bundle = world.entity_mut(e2).take::<(B, TableStored)>().unwrap();
assert_eq!(removed_bundle, (B(2), TableStored("2")));
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
let results = query
.iter(&world)
.map(|(a, b)| (a.0, b.0))
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
.collect::<Vec<_>>();
assert_eq!(results, vec![(1, "1"), (3, "3"),]);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
let mut a_query = world.query::<&A>();
let results = a_query.iter(&world).map(|a| a.0).collect::<Vec<_>>();
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
assert_eq!(results, vec![1, 3, 2]);
let entity_ref = world.entity(e2);
assert_eq!(
entity_ref.get::<A>(),
Some(&A(2)),
"A is not in the removed bundle, so it should exist"
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
);
assert_eq!(
entity_ref.get::<B>(),
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
None,
"B is in the removed bundle, so it should not exist"
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
);
assert_eq!(
entity_ref.get::<TableStored>(),
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
None,
"TableStored is in the removed bundle, so it should not exist"
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
);
}
#[test]
fn non_send_resource() {
let mut world = World::default();
world.insert_non_send_resource(123i32);
world.insert_non_send_resource(456i64);
assert_eq!(*world.non_send_resource::<i32>(), 123);
assert_eq!(*world.non_send_resource_mut::<i64>(), 456);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
}
Panic on dropping NonSend in non-origin thread. (#6534) # Objective Fixes #3310. Fixes #6282. Fixes #6278. Fixes #3666. ## Solution Split out `!Send` resources into `NonSendResources`. Add a `origin_thread_id` to all `!Send` Resources, check it on dropping `NonSendResourceData`, if there's a mismatch, panic. Moved all of the checks that `MainThreadValidator` would do into `NonSendResources` instead. All `!Send` resources now individually track which thread they were inserted from. This is validated against for every access, mutation, and drop that could be done against the value. A regression test using an altered version of the example from #3310 has been added. This is a stopgap solution for the current status quo. A full solution may involve fully removing `!Send` resources/components from `World`, which will likely require a much more thorough design on how to handle the existing in-engine and ecosystem use cases. This PR also introduces another breaking change: ```rust use bevy_ecs::prelude::*; #[derive(Resource)] struct Resource(u32); fn main() { let mut world = World::new(); world.insert_resource(Resource(1)); world.insert_non_send_resource(Resource(2)); let res = world.get_resource_mut::<Resource>().unwrap(); assert_eq!(res.0, 2); } ``` This code will run correctly on 0.9.1 but not with this PR, since NonSend resources and normal resources have become actual distinct concepts storage wise. ## Changelog Changed: Fix soundness bug with `World: Send`. Dropping a `World` that contains a `!Send` resource on the wrong thread will now panic. ## Migration Guide Normal resources and `NonSend` resources no longer share the same backing storage. If `R: Resource`, then `NonSend<R>` and `Res<R>` will return different instances from each other. If you are using both `Res<T>` and `NonSend<T>` (or their mutable variants), to fetch the same resources, it's strongly advised to use `Res<T>`.
2023-01-09 20:40:34 +00:00
#[test]
fn non_send_resource_points_to_distinct_data() {
let mut world = World::default();
world.insert_resource(A(123));
world.insert_non_send_resource(A(456));
assert_eq!(*world.resource::<A>(), A(123));
assert_eq!(*world.non_send_resource::<A>(), A(456));
}
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
#[test]
#[should_panic]
fn non_send_resource_panic() {
let mut world = World::default();
world.insert_non_send_resource(0i32);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
std::thread::spawn(move || {
let _ = world.non_send_resource_mut::<i32>();
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
})
.join()
.unwrap();
}
#[test]
fn exact_size_query() {
let mut world = World::default();
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
world.spawn((A(0), B(0)));
world.spawn((A(0), B(0)));
world.spawn((A(0), B(0), C));
world.spawn(C);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
let mut query = world.query::<(&A, &B)>();
assert_eq!(query.iter(&world).len(), 3);
}
#[test]
#[should_panic]
fn duplicate_components_panic() {
let mut world = World::new();
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
world.spawn((A(1), A(2)));
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
}
#[test]
#[should_panic]
fn ref_and_mut_query_panic() {
let mut world = World::new();
world.query::<(&A, &mut A)>();
}
Implement WorldQuery for EntityRef (#6960) # Objective Partially address #5504. Fix #4278. Provide "whole entity" access in queries. This can be useful when you don't know at compile time what you're accessing (i.e. reflection via `ReflectComponent`). ## Solution Implement `WorldQuery` for `EntityRef`. - This provides read-only access to the entire entity, and supports anything that `EntityRef` can normally do. - It matches all archetypes and tables and will densely iterate when possible. - It marks all of the ArchetypeComponentIds of a matched archetype as read. - Adding it to a query will cause it to panic if used in conjunction with any other mutable access. - Expanded the docs on Query to advertise this feature. - Added tests to ensure the panics were working as intended. - Added `EntityRef` to the ECS prelude. To make this safe, `EntityRef::world` was removed as it gave potential `UnsafeCell`-like access to other parts of the `World` including aliased mutable access to the components it would otherwise read safely. ## Performance Not great beyond the additional parallelization opportunity over exclusive systems. The `EntityRef` is fetched from `Entities` like any other call to `World::entity`, which can be very random access heavy. This could be simplified if `ArchetypeRow` is available in `WorldQuery::fetch`'s arguments, but that's likely not something we should optimize for. ## Future work An equivalent API where it gives mutable access to all components on a entity can be done with a scoped version of `EntityMut` where it does not provide `&mut World` access nor allow for structural changes to the entity is feasible as well. This could be done as a safe alternative to exclusive system when structural mutation isn't required or the target set of entities is scoped. --- ## Changelog Added: `Access::has_any_write` Added: `EntityRef` now implements `WorldQuery`. Allows read-only access to the entire entity, incompatible with any other mutable access, can be mixed with `With`/`Without` filters for more targeted use. Added: `EntityRef` to `bevy::ecs::prelude`. Removed: `EntityRef::world` ## Migration Guide TODO --------- Co-authored-by: Carter Weinberg <weinbergcarter@gmail.com> Co-authored-by: Jakob Hellermann <jakob.hellermann@protonmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2023-06-22 21:20:00 +00:00
#[test]
#[should_panic]
fn entity_ref_and_mut_query_panic() {
let mut world = World::new();
world.query::<(EntityRef, &mut A)>();
}
#[test]
#[should_panic]
fn mut_and_ref_query_panic() {
let mut world = World::new();
world.query::<(&mut A, &A)>();
}
Implement WorldQuery for EntityRef (#6960) # Objective Partially address #5504. Fix #4278. Provide "whole entity" access in queries. This can be useful when you don't know at compile time what you're accessing (i.e. reflection via `ReflectComponent`). ## Solution Implement `WorldQuery` for `EntityRef`. - This provides read-only access to the entire entity, and supports anything that `EntityRef` can normally do. - It matches all archetypes and tables and will densely iterate when possible. - It marks all of the ArchetypeComponentIds of a matched archetype as read. - Adding it to a query will cause it to panic if used in conjunction with any other mutable access. - Expanded the docs on Query to advertise this feature. - Added tests to ensure the panics were working as intended. - Added `EntityRef` to the ECS prelude. To make this safe, `EntityRef::world` was removed as it gave potential `UnsafeCell`-like access to other parts of the `World` including aliased mutable access to the components it would otherwise read safely. ## Performance Not great beyond the additional parallelization opportunity over exclusive systems. The `EntityRef` is fetched from `Entities` like any other call to `World::entity`, which can be very random access heavy. This could be simplified if `ArchetypeRow` is available in `WorldQuery::fetch`'s arguments, but that's likely not something we should optimize for. ## Future work An equivalent API where it gives mutable access to all components on a entity can be done with a scoped version of `EntityMut` where it does not provide `&mut World` access nor allow for structural changes to the entity is feasible as well. This could be done as a safe alternative to exclusive system when structural mutation isn't required or the target set of entities is scoped. --- ## Changelog Added: `Access::has_any_write` Added: `EntityRef` now implements `WorldQuery`. Allows read-only access to the entire entity, incompatible with any other mutable access, can be mixed with `With`/`Without` filters for more targeted use. Added: `EntityRef` to `bevy::ecs::prelude`. Removed: `EntityRef::world` ## Migration Guide TODO --------- Co-authored-by: Carter Weinberg <weinbergcarter@gmail.com> Co-authored-by: Jakob Hellermann <jakob.hellermann@protonmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2023-06-22 21:20:00 +00:00
#[test]
#[should_panic]
fn mut_and_entity_ref_query_panic() {
let mut world = World::new();
world.query::<(&mut A, EntityRef)>();
}
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
#[test]
#[should_panic]
fn mut_and_mut_query_panic() {
let mut world = World::new();
world.query::<(&mut A, &mut A)>();
}
#[test]
#[should_panic]
fn multiple_worlds_same_query_iter() {
let mut world_a = World::new();
let world_b = World::new();
let mut query = world_a.query::<&A>();
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
query.iter(&world_a);
query.iter(&world_b);
}
#[test]
fn query_filters_dont_collide_with_fetches() {
let mut world = World::new();
world.query_filtered::<&mut A, Changed<A>>();
}
#[test]
fn filtered_query_access() {
let mut world = World::new();
let query = world.query_filtered::<&mut A, Changed<B>>();
let mut expected = FilteredAccess::<ComponentId>::default();
let a_id = world.components.get_id(TypeId::of::<A>()).unwrap();
let b_id = world.components.get_id(TypeId::of::<B>()).unwrap();
expected.add_write(a_id);
expected.add_read(b_id);
assert!(
query.component_access.eq(&expected),
"ComponentId access from query fetch and query filter should be combined"
);
}
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
#[test]
#[should_panic]
fn multiple_worlds_same_query_get() {
let mut world_a = World::new();
let world_b = World::new();
let mut query = world_a.query::<&A>();
let _ = query.get(&world_a, Entity::from_raw(0));
let _ = query.get(&world_b, Entity::from_raw(0));
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
}
#[test]
#[should_panic]
fn multiple_worlds_same_query_for_each() {
let mut world_a = World::new();
let world_b = World::new();
let mut query = world_a.query::<&A>();
Override QueryIter::fold to port Query::for_each perf gains to select Iterator combinators (#6773) # Objective After #6547, `Query::for_each` has been capable of automatic vectorization on certain queries, which is seeing a notable (>50% CPU time improvements) for iteration. However, `Query::for_each` isn't idiomatic Rust, and lacks the flexibility of iterator combinators. Ideally, `Query::iter` and friends should be able to achieve the same results. However, this does seem to blocked upstream (rust-lang/rust#104914) by Rust's loop optimizations. ## Solution This is an intermediate solution and refactor. This moves the `Query::for_each` implementation onto the `Iterator::fold` implementation for `QueryIter` instead. This should result in the same automatic vectorization optimization on all `Iterator` functions that internally use fold, including `Iterator::for_each`, `Iterator::count`, etc. With this, it should close the gap between the two completely. Internally, this PR changes `Query::for_each` to use `query.iter().for_each(..)` instead of the duplicated implementation. Separately, the duplicate implementations of internal iteration (i.e. `Query::par_for_each`) now use portions of the current `Query::for_each` implementation factored out into their own functions. This also massively cleans up our internal fragmentation of internal iteration options, deduplicating the iteration code used in `for_each` and `par_iter().for_each()`. --- ## Changelog Changed: `Query::for_each`, `Query::for_each_mut`, `Query::for_each`, and `Query::for_each_mut` have been moved to `QueryIter`'s `Iterator::for_each` implementation, and still retains their performance improvements over normal iteration. These APIs are deprecated in 0.13 and will be removed in 0.14. --------- Co-authored-by: JoJoJet <21144246+JoJoJet@users.noreply.github.com> Co-authored-by: Alice Cecile <alice.i.cecile@gmail.com>
2023-12-01 09:09:55 +00:00
query.iter(&world_a).for_each(|_| {});
query.iter(&world_b).for_each(|_| {});
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
}
#[test]
fn resource_scope() {
let mut world = World::default();
Make `Resource` trait opt-in, requiring `#[derive(Resource)]` V2 (#5577) *This PR description is an edited copy of #5007, written by @alice-i-cecile.* # Objective Follow-up to https://github.com/bevyengine/bevy/pull/2254. The `Resource` trait currently has a blanket implementation for all types that meet its bounds. While ergonomic, this results in several drawbacks: * it is possible to make confusing, silent mistakes such as inserting a function pointer (Foo) rather than a value (Foo::Bar) as a resource * it is challenging to discover if a type is intended to be used as a resource * we cannot later add customization options (see the [RFC](https://github.com/bevyengine/rfcs/blob/main/rfcs/27-derive-component.md) for the equivalent choice for Component). * dependencies can use the same Rust type as a resource in invisibly conflicting ways * raw Rust types used as resources cannot preserve privacy appropriately, as anyone able to access that type can read and write to internal values * we cannot capture a definitive list of possible resources to display to users in an editor ## Notes to reviewers * Review this commit-by-commit; there's effectively no back-tracking and there's a lot of churn in some of these commits. *ira: My commits are not as well organized :')* * I've relaxed the bound on Local to Send + Sync + 'static: I don't think these concerns apply there, so this can keep things simple. Storing e.g. a u32 in a Local is fine, because there's a variable name attached explaining what it does. * I think this is a bad place for the Resource trait to live, but I've left it in place to make reviewing easier. IMO that's best tackled with https://github.com/bevyengine/bevy/issues/4981. ## Changelog `Resource` is no longer automatically implemented for all matching types. Instead, use the new `#[derive(Resource)]` macro. ## Migration Guide Add `#[derive(Resource)]` to all types you are using as a resource. If you are using a third party type as a resource, wrap it in a tuple struct to bypass orphan rules. Consider deriving `Deref` and `DerefMut` to improve ergonomics. `ClearColor` no longer implements `Component`. Using `ClearColor` as a component in 0.8 did nothing. Use the `ClearColorConfig` in the `Camera3d` and `Camera2d` components instead. Co-authored-by: Alice <alice.i.cecile@gmail.com> Co-authored-by: Alice Cecile <alice.i.cecile@gmail.com> Co-authored-by: devil-ira <justthecooldude@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-08-08 21:36:35 +00:00
world.insert_resource(A(0));
world.resource_scope(|world: &mut World, mut value: Mut<A>| {
value.0 += 1;
assert!(!world.contains_resource::<A>());
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
});
Make `Resource` trait opt-in, requiring `#[derive(Resource)]` V2 (#5577) *This PR description is an edited copy of #5007, written by @alice-i-cecile.* # Objective Follow-up to https://github.com/bevyengine/bevy/pull/2254. The `Resource` trait currently has a blanket implementation for all types that meet its bounds. While ergonomic, this results in several drawbacks: * it is possible to make confusing, silent mistakes such as inserting a function pointer (Foo) rather than a value (Foo::Bar) as a resource * it is challenging to discover if a type is intended to be used as a resource * we cannot later add customization options (see the [RFC](https://github.com/bevyengine/rfcs/blob/main/rfcs/27-derive-component.md) for the equivalent choice for Component). * dependencies can use the same Rust type as a resource in invisibly conflicting ways * raw Rust types used as resources cannot preserve privacy appropriately, as anyone able to access that type can read and write to internal values * we cannot capture a definitive list of possible resources to display to users in an editor ## Notes to reviewers * Review this commit-by-commit; there's effectively no back-tracking and there's a lot of churn in some of these commits. *ira: My commits are not as well organized :')* * I've relaxed the bound on Local to Send + Sync + 'static: I don't think these concerns apply there, so this can keep things simple. Storing e.g. a u32 in a Local is fine, because there's a variable name attached explaining what it does. * I think this is a bad place for the Resource trait to live, but I've left it in place to make reviewing easier. IMO that's best tackled with https://github.com/bevyengine/bevy/issues/4981. ## Changelog `Resource` is no longer automatically implemented for all matching types. Instead, use the new `#[derive(Resource)]` macro. ## Migration Guide Add `#[derive(Resource)]` to all types you are using as a resource. If you are using a third party type as a resource, wrap it in a tuple struct to bypass orphan rules. Consider deriving `Deref` and `DerefMut` to improve ergonomics. `ClearColor` no longer implements `Component`. Using `ClearColor` as a component in 0.8 did nothing. Use the `ClearColorConfig` in the `Camera3d` and `Camera2d` components instead. Co-authored-by: Alice <alice.i.cecile@gmail.com> Co-authored-by: Alice Cecile <alice.i.cecile@gmail.com> Co-authored-by: devil-ira <justthecooldude@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-08-08 21:36:35 +00:00
assert_eq!(world.resource::<A>().0, 1);
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
}
#[test]
#[should_panic(
Panic on dropping NonSend in non-origin thread. (#6534) # Objective Fixes #3310. Fixes #6282. Fixes #6278. Fixes #3666. ## Solution Split out `!Send` resources into `NonSendResources`. Add a `origin_thread_id` to all `!Send` Resources, check it on dropping `NonSendResourceData`, if there's a mismatch, panic. Moved all of the checks that `MainThreadValidator` would do into `NonSendResources` instead. All `!Send` resources now individually track which thread they were inserted from. This is validated against for every access, mutation, and drop that could be done against the value. A regression test using an altered version of the example from #3310 has been added. This is a stopgap solution for the current status quo. A full solution may involve fully removing `!Send` resources/components from `World`, which will likely require a much more thorough design on how to handle the existing in-engine and ecosystem use cases. This PR also introduces another breaking change: ```rust use bevy_ecs::prelude::*; #[derive(Resource)] struct Resource(u32); fn main() { let mut world = World::new(); world.insert_resource(Resource(1)); world.insert_non_send_resource(Resource(2)); let res = world.get_resource_mut::<Resource>().unwrap(); assert_eq!(res.0, 2); } ``` This code will run correctly on 0.9.1 but not with this PR, since NonSend resources and normal resources have become actual distinct concepts storage wise. ## Changelog Changed: Fix soundness bug with `World: Send`. Dropping a `World` that contains a `!Send` resource on the wrong thread will now panic. ## Migration Guide Normal resources and `NonSend` resources no longer share the same backing storage. If `R: Resource`, then `NonSend<R>` and `Res<R>` will return different instances from each other. If you are using both `Res<T>` and `NonSend<T>` (or their mutable variants), to fetch the same resources, it's strongly advised to use `Res<T>`.
2023-01-09 20:40:34 +00:00
expected = "Attempted to access or drop non-send resource bevy_ecs::tests::NonSendA from thread"
)]
Panic on dropping NonSend in non-origin thread. (#6534) # Objective Fixes #3310. Fixes #6282. Fixes #6278. Fixes #3666. ## Solution Split out `!Send` resources into `NonSendResources`. Add a `origin_thread_id` to all `!Send` Resources, check it on dropping `NonSendResourceData`, if there's a mismatch, panic. Moved all of the checks that `MainThreadValidator` would do into `NonSendResources` instead. All `!Send` resources now individually track which thread they were inserted from. This is validated against for every access, mutation, and drop that could be done against the value. A regression test using an altered version of the example from #3310 has been added. This is a stopgap solution for the current status quo. A full solution may involve fully removing `!Send` resources/components from `World`, which will likely require a much more thorough design on how to handle the existing in-engine and ecosystem use cases. This PR also introduces another breaking change: ```rust use bevy_ecs::prelude::*; #[derive(Resource)] struct Resource(u32); fn main() { let mut world = World::new(); world.insert_resource(Resource(1)); world.insert_non_send_resource(Resource(2)); let res = world.get_resource_mut::<Resource>().unwrap(); assert_eq!(res.0, 2); } ``` This code will run correctly on 0.9.1 but not with this PR, since NonSend resources and normal resources have become actual distinct concepts storage wise. ## Changelog Changed: Fix soundness bug with `World: Send`. Dropping a `World` that contains a `!Send` resource on the wrong thread will now panic. ## Migration Guide Normal resources and `NonSend` resources no longer share the same backing storage. If `R: Resource`, then `NonSend<R>` and `Res<R>` will return different instances from each other. If you are using both `Res<T>` and `NonSend<T>` (or their mutable variants), to fetch the same resources, it's strongly advised to use `Res<T>`.
2023-01-09 20:40:34 +00:00
fn non_send_resource_drop_from_different_thread() {
let mut world = World::default();
world.insert_non_send_resource(NonSendA::default());
let thread = std::thread::spawn(move || {
Panic on dropping NonSend in non-origin thread. (#6534) # Objective Fixes #3310. Fixes #6282. Fixes #6278. Fixes #3666. ## Solution Split out `!Send` resources into `NonSendResources`. Add a `origin_thread_id` to all `!Send` Resources, check it on dropping `NonSendResourceData`, if there's a mismatch, panic. Moved all of the checks that `MainThreadValidator` would do into `NonSendResources` instead. All `!Send` resources now individually track which thread they were inserted from. This is validated against for every access, mutation, and drop that could be done against the value. A regression test using an altered version of the example from #3310 has been added. This is a stopgap solution for the current status quo. A full solution may involve fully removing `!Send` resources/components from `World`, which will likely require a much more thorough design on how to handle the existing in-engine and ecosystem use cases. This PR also introduces another breaking change: ```rust use bevy_ecs::prelude::*; #[derive(Resource)] struct Resource(u32); fn main() { let mut world = World::new(); world.insert_resource(Resource(1)); world.insert_non_send_resource(Resource(2)); let res = world.get_resource_mut::<Resource>().unwrap(); assert_eq!(res.0, 2); } ``` This code will run correctly on 0.9.1 but not with this PR, since NonSend resources and normal resources have become actual distinct concepts storage wise. ## Changelog Changed: Fix soundness bug with `World: Send`. Dropping a `World` that contains a `!Send` resource on the wrong thread will now panic. ## Migration Guide Normal resources and `NonSend` resources no longer share the same backing storage. If `R: Resource`, then `NonSend<R>` and `Res<R>` will return different instances from each other. If you are using both `Res<T>` and `NonSend<T>` (or their mutable variants), to fetch the same resources, it's strongly advised to use `Res<T>`.
2023-01-09 20:40:34 +00:00
// Dropping the non-send resource on a different thread
// Should result in a panic
Panic on dropping NonSend in non-origin thread. (#6534) # Objective Fixes #3310. Fixes #6282. Fixes #6278. Fixes #3666. ## Solution Split out `!Send` resources into `NonSendResources`. Add a `origin_thread_id` to all `!Send` Resources, check it on dropping `NonSendResourceData`, if there's a mismatch, panic. Moved all of the checks that `MainThreadValidator` would do into `NonSendResources` instead. All `!Send` resources now individually track which thread they were inserted from. This is validated against for every access, mutation, and drop that could be done against the value. A regression test using an altered version of the example from #3310 has been added. This is a stopgap solution for the current status quo. A full solution may involve fully removing `!Send` resources/components from `World`, which will likely require a much more thorough design on how to handle the existing in-engine and ecosystem use cases. This PR also introduces another breaking change: ```rust use bevy_ecs::prelude::*; #[derive(Resource)] struct Resource(u32); fn main() { let mut world = World::new(); world.insert_resource(Resource(1)); world.insert_non_send_resource(Resource(2)); let res = world.get_resource_mut::<Resource>().unwrap(); assert_eq!(res.0, 2); } ``` This code will run correctly on 0.9.1 but not with this PR, since NonSend resources and normal resources have become actual distinct concepts storage wise. ## Changelog Changed: Fix soundness bug with `World: Send`. Dropping a `World` that contains a `!Send` resource on the wrong thread will now panic. ## Migration Guide Normal resources and `NonSend` resources no longer share the same backing storage. If `R: Resource`, then `NonSend<R>` and `Res<R>` will return different instances from each other. If you are using both `Res<T>` and `NonSend<T>` (or their mutable variants), to fetch the same resources, it's strongly advised to use `Res<T>`.
2023-01-09 20:40:34 +00:00
drop(world);
});
if let Err(err) = thread.join() {
std::panic::resume_unwind(err);
}
}
Panic on dropping NonSend in non-origin thread. (#6534) # Objective Fixes #3310. Fixes #6282. Fixes #6278. Fixes #3666. ## Solution Split out `!Send` resources into `NonSendResources`. Add a `origin_thread_id` to all `!Send` Resources, check it on dropping `NonSendResourceData`, if there's a mismatch, panic. Moved all of the checks that `MainThreadValidator` would do into `NonSendResources` instead. All `!Send` resources now individually track which thread they were inserted from. This is validated against for every access, mutation, and drop that could be done against the value. A regression test using an altered version of the example from #3310 has been added. This is a stopgap solution for the current status quo. A full solution may involve fully removing `!Send` resources/components from `World`, which will likely require a much more thorough design on how to handle the existing in-engine and ecosystem use cases. This PR also introduces another breaking change: ```rust use bevy_ecs::prelude::*; #[derive(Resource)] struct Resource(u32); fn main() { let mut world = World::new(); world.insert_resource(Resource(1)); world.insert_non_send_resource(Resource(2)); let res = world.get_resource_mut::<Resource>().unwrap(); assert_eq!(res.0, 2); } ``` This code will run correctly on 0.9.1 but not with this PR, since NonSend resources and normal resources have become actual distinct concepts storage wise. ## Changelog Changed: Fix soundness bug with `World: Send`. Dropping a `World` that contains a `!Send` resource on the wrong thread will now panic. ## Migration Guide Normal resources and `NonSend` resources no longer share the same backing storage. If `R: Resource`, then `NonSend<R>` and `Res<R>` will return different instances from each other. If you are using both `Res<T>` and `NonSend<T>` (or their mutable variants), to fetch the same resources, it's strongly advised to use `Res<T>`.
2023-01-09 20:40:34 +00:00
#[test]
fn non_send_resource_drop_from_same_thread() {
let mut world = World::default();
world.insert_non_send_resource(NonSendA::default());
drop(world);
}
#[test]
fn insert_overwrite_drop() {
let (dropck1, dropped1) = DropCk::new_pair();
let (dropck2, dropped2) = DropCk::new_pair();
let mut world = World::default();
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
world.spawn(dropck1).insert(dropck2);
assert_eq!(dropped1.load(Ordering::Relaxed), 1);
assert_eq!(dropped2.load(Ordering::Relaxed), 0);
drop(world);
assert_eq!(dropped1.load(Ordering::Relaxed), 1);
assert_eq!(dropped2.load(Ordering::Relaxed), 1);
}
#[test]
fn insert_overwrite_drop_sparse() {
let (dropck1, dropped1) = DropCk::new_pair();
let (dropck2, dropped2) = DropCk::new_pair();
let mut world = World::default();
world
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
.spawn(DropCkSparse(dropck1))
.insert(DropCkSparse(dropck2));
assert_eq!(dropped1.load(Ordering::Relaxed), 1);
assert_eq!(dropped2.load(Ordering::Relaxed), 0);
drop(world);
assert_eq!(dropped1.load(Ordering::Relaxed), 1);
assert_eq!(dropped2.load(Ordering::Relaxed), 1);
}
Spawn specific entities: spawn or insert operations, refactor spawn internals, world clearing (#2673) This upstreams the code changes used by the new renderer to enable cross-app Entity reuse: * Spawning at specific entities * get_or_spawn: spawns an entity if it doesn't already exist and returns an EntityMut * insert_or_spawn_batch: the batched equivalent to `world.get_or_spawn(entity).insert_bundle(bundle)` * Clearing entities and storages * Allocating Entities with "invalid" archetypes. These entities cannot be queried / are treated as "non existent". They serve as "reserved" entities that won't show up when calling `spawn()`. They must be "specifically spawned at" using apis like `get_or_spawn(entity)`. In combination, these changes enable the "render world" to clear entities / storages each frame and reserve all "app world entities". These can then be spawned during the "render extract step". This refactors "spawn" and "insert" code in a way that I think is a massive improvement to legibility and re-usability. It also yields marginal performance wins by reducing some duplicate lookups (less than a percentage point improvement on insertion benchmarks). There is also some potential for future unsafe reduction (by making BatchSpawner and BatchInserter generic). But for now I want to cut down generic usage to a minimum to encourage smaller binaries and faster compiles. This is currently a draft because it needs more tests (although this code has already had some real-world testing on my custom-shaders branch). I also fixed the benchmarks (which currently don't compile!) / added new ones to illustrate batching wins. After these changes, Bevy ECS is basically ready to accommodate the new renderer. I think the biggest missing piece at this point is "sub apps".
2021-08-25 23:34:02 +00:00
#[test]
fn clear_entities() {
let mut world = World::default();
Make `Resource` trait opt-in, requiring `#[derive(Resource)]` V2 (#5577) *This PR description is an edited copy of #5007, written by @alice-i-cecile.* # Objective Follow-up to https://github.com/bevyengine/bevy/pull/2254. The `Resource` trait currently has a blanket implementation for all types that meet its bounds. While ergonomic, this results in several drawbacks: * it is possible to make confusing, silent mistakes such as inserting a function pointer (Foo) rather than a value (Foo::Bar) as a resource * it is challenging to discover if a type is intended to be used as a resource * we cannot later add customization options (see the [RFC](https://github.com/bevyengine/rfcs/blob/main/rfcs/27-derive-component.md) for the equivalent choice for Component). * dependencies can use the same Rust type as a resource in invisibly conflicting ways * raw Rust types used as resources cannot preserve privacy appropriately, as anyone able to access that type can read and write to internal values * we cannot capture a definitive list of possible resources to display to users in an editor ## Notes to reviewers * Review this commit-by-commit; there's effectively no back-tracking and there's a lot of churn in some of these commits. *ira: My commits are not as well organized :')* * I've relaxed the bound on Local to Send + Sync + 'static: I don't think these concerns apply there, so this can keep things simple. Storing e.g. a u32 in a Local is fine, because there's a variable name attached explaining what it does. * I think this is a bad place for the Resource trait to live, but I've left it in place to make reviewing easier. IMO that's best tackled with https://github.com/bevyengine/bevy/issues/4981. ## Changelog `Resource` is no longer automatically implemented for all matching types. Instead, use the new `#[derive(Resource)]` macro. ## Migration Guide Add `#[derive(Resource)]` to all types you are using as a resource. If you are using a third party type as a resource, wrap it in a tuple struct to bypass orphan rules. Consider deriving `Deref` and `DerefMut` to improve ergonomics. `ClearColor` no longer implements `Component`. Using `ClearColor` as a component in 0.8 did nothing. Use the `ClearColorConfig` in the `Camera3d` and `Camera2d` components instead. Co-authored-by: Alice <alice.i.cecile@gmail.com> Co-authored-by: Alice Cecile <alice.i.cecile@gmail.com> Co-authored-by: devil-ira <justthecooldude@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-08-08 21:36:35 +00:00
world.insert_resource(A(0));
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
world.spawn(A(1));
world.spawn(SparseStored(1));
Spawn specific entities: spawn or insert operations, refactor spawn internals, world clearing (#2673) This upstreams the code changes used by the new renderer to enable cross-app Entity reuse: * Spawning at specific entities * get_or_spawn: spawns an entity if it doesn't already exist and returns an EntityMut * insert_or_spawn_batch: the batched equivalent to `world.get_or_spawn(entity).insert_bundle(bundle)` * Clearing entities and storages * Allocating Entities with "invalid" archetypes. These entities cannot be queried / are treated as "non existent". They serve as "reserved" entities that won't show up when calling `spawn()`. They must be "specifically spawned at" using apis like `get_or_spawn(entity)`. In combination, these changes enable the "render world" to clear entities / storages each frame and reserve all "app world entities". These can then be spawned during the "render extract step". This refactors "spawn" and "insert" code in a way that I think is a massive improvement to legibility and re-usability. It also yields marginal performance wins by reducing some duplicate lookups (less than a percentage point improvement on insertion benchmarks). There is also some potential for future unsafe reduction (by making BatchSpawner and BatchInserter generic). But for now I want to cut down generic usage to a minimum to encourage smaller binaries and faster compiles. This is currently a draft because it needs more tests (although this code has already had some real-world testing on my custom-shaders branch). I also fixed the benchmarks (which currently don't compile!) / added new ones to illustrate batching wins. After these changes, Bevy ECS is basically ready to accommodate the new renderer. I think the biggest missing piece at this point is "sub apps".
2021-08-25 23:34:02 +00:00
let mut q1 = world.query::<&A>();
let mut q2 = world.query::<&SparseStored>();
Spawn specific entities: spawn or insert operations, refactor spawn internals, world clearing (#2673) This upstreams the code changes used by the new renderer to enable cross-app Entity reuse: * Spawning at specific entities * get_or_spawn: spawns an entity if it doesn't already exist and returns an EntityMut * insert_or_spawn_batch: the batched equivalent to `world.get_or_spawn(entity).insert_bundle(bundle)` * Clearing entities and storages * Allocating Entities with "invalid" archetypes. These entities cannot be queried / are treated as "non existent". They serve as "reserved" entities that won't show up when calling `spawn()`. They must be "specifically spawned at" using apis like `get_or_spawn(entity)`. In combination, these changes enable the "render world" to clear entities / storages each frame and reserve all "app world entities". These can then be spawned during the "render extract step". This refactors "spawn" and "insert" code in a way that I think is a massive improvement to legibility and re-usability. It also yields marginal performance wins by reducing some duplicate lookups (less than a percentage point improvement on insertion benchmarks). There is also some potential for future unsafe reduction (by making BatchSpawner and BatchInserter generic). But for now I want to cut down generic usage to a minimum to encourage smaller binaries and faster compiles. This is currently a draft because it needs more tests (although this code has already had some real-world testing on my custom-shaders branch). I also fixed the benchmarks (which currently don't compile!) / added new ones to illustrate batching wins. After these changes, Bevy ECS is basically ready to accommodate the new renderer. I think the biggest missing piece at this point is "sub apps".
2021-08-25 23:34:02 +00:00
assert_eq!(q1.iter(&world).len(), 1);
assert_eq!(q2.iter(&world).len(), 1);
assert_eq!(world.entities().len(), 2);
world.clear_entities();
assert_eq!(
q1.iter(&world).len(),
0,
"world should not contain table components"
);
assert_eq!(
q2.iter(&world).len(),
0,
"world should not contain sparse set components"
);
assert_eq!(
world.entities().len(),
0,
"world should not have any entities"
);
assert_eq!(
Make `Resource` trait opt-in, requiring `#[derive(Resource)]` V2 (#5577) *This PR description is an edited copy of #5007, written by @alice-i-cecile.* # Objective Follow-up to https://github.com/bevyengine/bevy/pull/2254. The `Resource` trait currently has a blanket implementation for all types that meet its bounds. While ergonomic, this results in several drawbacks: * it is possible to make confusing, silent mistakes such as inserting a function pointer (Foo) rather than a value (Foo::Bar) as a resource * it is challenging to discover if a type is intended to be used as a resource * we cannot later add customization options (see the [RFC](https://github.com/bevyengine/rfcs/blob/main/rfcs/27-derive-component.md) for the equivalent choice for Component). * dependencies can use the same Rust type as a resource in invisibly conflicting ways * raw Rust types used as resources cannot preserve privacy appropriately, as anyone able to access that type can read and write to internal values * we cannot capture a definitive list of possible resources to display to users in an editor ## Notes to reviewers * Review this commit-by-commit; there's effectively no back-tracking and there's a lot of churn in some of these commits. *ira: My commits are not as well organized :')* * I've relaxed the bound on Local to Send + Sync + 'static: I don't think these concerns apply there, so this can keep things simple. Storing e.g. a u32 in a Local is fine, because there's a variable name attached explaining what it does. * I think this is a bad place for the Resource trait to live, but I've left it in place to make reviewing easier. IMO that's best tackled with https://github.com/bevyengine/bevy/issues/4981. ## Changelog `Resource` is no longer automatically implemented for all matching types. Instead, use the new `#[derive(Resource)]` macro. ## Migration Guide Add `#[derive(Resource)]` to all types you are using as a resource. If you are using a third party type as a resource, wrap it in a tuple struct to bypass orphan rules. Consider deriving `Deref` and `DerefMut` to improve ergonomics. `ClearColor` no longer implements `Component`. Using `ClearColor` as a component in 0.8 did nothing. Use the `ClearColorConfig` in the `Camera3d` and `Camera2d` components instead. Co-authored-by: Alice <alice.i.cecile@gmail.com> Co-authored-by: Alice Cecile <alice.i.cecile@gmail.com> Co-authored-by: devil-ira <justthecooldude@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-08-08 21:36:35 +00:00
world.resource::<A>().0,
Spawn specific entities: spawn or insert operations, refactor spawn internals, world clearing (#2673) This upstreams the code changes used by the new renderer to enable cross-app Entity reuse: * Spawning at specific entities * get_or_spawn: spawns an entity if it doesn't already exist and returns an EntityMut * insert_or_spawn_batch: the batched equivalent to `world.get_or_spawn(entity).insert_bundle(bundle)` * Clearing entities and storages * Allocating Entities with "invalid" archetypes. These entities cannot be queried / are treated as "non existent". They serve as "reserved" entities that won't show up when calling `spawn()`. They must be "specifically spawned at" using apis like `get_or_spawn(entity)`. In combination, these changes enable the "render world" to clear entities / storages each frame and reserve all "app world entities". These can then be spawned during the "render extract step". This refactors "spawn" and "insert" code in a way that I think is a massive improvement to legibility and re-usability. It also yields marginal performance wins by reducing some duplicate lookups (less than a percentage point improvement on insertion benchmarks). There is also some potential for future unsafe reduction (by making BatchSpawner and BatchInserter generic). But for now I want to cut down generic usage to a minimum to encourage smaller binaries and faster compiles. This is currently a draft because it needs more tests (although this code has already had some real-world testing on my custom-shaders branch). I also fixed the benchmarks (which currently don't compile!) / added new ones to illustrate batching wins. After these changes, Bevy ECS is basically ready to accommodate the new renderer. I think the biggest missing piece at this point is "sub apps".
2021-08-25 23:34:02 +00:00
0,
"world should still contain resources"
);
}
#[test]
fn test_is_archetypal_size_hints() {
let mut world = World::default();
macro_rules! query_min_size {
($query:ty, $filter:ty) => {
world
.query_filtered::<$query, $filter>()
.iter(&world)
.size_hint()
.0
};
}
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
world.spawn((A(1), B(1), C));
world.spawn((A(1), C));
world.spawn((A(1), B(1)));
world.spawn((B(1), C));
world.spawn(A(1));
world.spawn(C);
assert_eq!(2, query_min_size![(), (With<A>, Without<B>)]);
assert_eq!(3, query_min_size![&B, Or<(With<A>, With<C>)>]);
assert_eq!(1, query_min_size![&B, (With<A>, With<C>)]);
assert_eq!(1, query_min_size![(&A, &B), With<C>]);
assert_eq!(4, query_min_size![&A, ()], "Simple Archetypal");
assert_eq!(4, query_min_size![Ref<A>, ()]);
// All the following should set minimum size to 0, as it's impossible to predict
// how many entities the filters will trim.
assert_eq!(0, query_min_size![(), Added<A>], "Simple Added");
assert_eq!(0, query_min_size![(), Changed<A>], "Simple Changed");
assert_eq!(0, query_min_size![(&A, &B), Changed<A>]);
assert_eq!(0, query_min_size![&A, (Changed<A>, With<B>)]);
assert_eq!(0, query_min_size![(&A, &B), Or<(Changed<A>, Changed<B>)>]);
}
Spawn specific entities: spawn or insert operations, refactor spawn internals, world clearing (#2673) This upstreams the code changes used by the new renderer to enable cross-app Entity reuse: * Spawning at specific entities * get_or_spawn: spawns an entity if it doesn't already exist and returns an EntityMut * insert_or_spawn_batch: the batched equivalent to `world.get_or_spawn(entity).insert_bundle(bundle)` * Clearing entities and storages * Allocating Entities with "invalid" archetypes. These entities cannot be queried / are treated as "non existent". They serve as "reserved" entities that won't show up when calling `spawn()`. They must be "specifically spawned at" using apis like `get_or_spawn(entity)`. In combination, these changes enable the "render world" to clear entities / storages each frame and reserve all "app world entities". These can then be spawned during the "render extract step". This refactors "spawn" and "insert" code in a way that I think is a massive improvement to legibility and re-usability. It also yields marginal performance wins by reducing some duplicate lookups (less than a percentage point improvement on insertion benchmarks). There is also some potential for future unsafe reduction (by making BatchSpawner and BatchInserter generic). But for now I want to cut down generic usage to a minimum to encourage smaller binaries and faster compiles. This is currently a draft because it needs more tests (although this code has already had some real-world testing on my custom-shaders branch). I also fixed the benchmarks (which currently don't compile!) / added new ones to illustrate batching wins. After these changes, Bevy ECS is basically ready to accommodate the new renderer. I think the biggest missing piece at this point is "sub apps".
2021-08-25 23:34:02 +00:00
#[test]
fn reserve_entities_across_worlds() {
let mut world_a = World::default();
let mut world_b = World::default();
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
let e1 = world_a.spawn(A(1)).id();
let e2 = world_a.spawn(A(2)).id();
Spawn specific entities: spawn or insert operations, refactor spawn internals, world clearing (#2673) This upstreams the code changes used by the new renderer to enable cross-app Entity reuse: * Spawning at specific entities * get_or_spawn: spawns an entity if it doesn't already exist and returns an EntityMut * insert_or_spawn_batch: the batched equivalent to `world.get_or_spawn(entity).insert_bundle(bundle)` * Clearing entities and storages * Allocating Entities with "invalid" archetypes. These entities cannot be queried / are treated as "non existent". They serve as "reserved" entities that won't show up when calling `spawn()`. They must be "specifically spawned at" using apis like `get_or_spawn(entity)`. In combination, these changes enable the "render world" to clear entities / storages each frame and reserve all "app world entities". These can then be spawned during the "render extract step". This refactors "spawn" and "insert" code in a way that I think is a massive improvement to legibility and re-usability. It also yields marginal performance wins by reducing some duplicate lookups (less than a percentage point improvement on insertion benchmarks). There is also some potential for future unsafe reduction (by making BatchSpawner and BatchInserter generic). But for now I want to cut down generic usage to a minimum to encourage smaller binaries and faster compiles. This is currently a draft because it needs more tests (although this code has already had some real-world testing on my custom-shaders branch). I also fixed the benchmarks (which currently don't compile!) / added new ones to illustrate batching wins. After these changes, Bevy ECS is basically ready to accommodate the new renderer. I think the biggest missing piece at this point is "sub apps".
2021-08-25 23:34:02 +00:00
let e3 = world_a.entities().reserve_entity();
world_a.flush();
Lock down access to Entities (#6740) # Objective The soundness of the ECS `World` partially relies on the correctness of the state of `Entities` stored within it. We're currently allowing users to (unsafely) mutate it, as well as readily construct it without using a `World`. While this is not strictly unsound so long as users (including `bevy_render`) safely use the APIs, it's a fairly easy path to unsoundness without much of a guard rail. Addresses #3362 for `bevy_ecs::entity`. Incorporates the changes from #3985. ## Solution Remove `Entities`'s `Default` implementation and force access to the type to only be through a properly constructed `World`. Additional cleanup for other parts of `bevy_ecs::entity`: - `Entity::index` and `Entity::generation` are no longer `pub(crate)`, opting to force the rest of bevy_ecs to use the public interface to access these values. - `EntityMeta` is no longer `pub` and also not `pub(crate)` to attempt to cut down on updating `generation` without going through an `Entities` API. It's currently inaccessible except via the `pub(crate)` Vec on `Entities`, there was no way for an outside user to use it. - Added `Entities::set`, an unsafe `pub(crate)` API for setting the location of an Entity (parallel to `Entities::get`) that replaces the internal case where we need to set the location of an entity when it's been spawned, moved, or despawned. - `Entities::alloc_at_without_replacement` is only used in `World::get_or_spawn` within the first party crates, and I cannot find a public use of this API in any ecosystem crate that I've checked (via GitHub search). - Attempted to document the few remaining undocumented public APIs in the module. --- ## Changelog Removed: `Entities`'s `Default` implementation. Removed: `EntityMeta` Removed: `Entities::alloc_at_without_replacement` and `AllocAtWithoutReplacement`. Co-authored-by: james7132 <contact@jamessliu.com> Co-authored-by: James Liu <contact@jamessliu.com>
2022-11-28 20:39:02 +00:00
let world_a_max_entities = world_a.entities().len();
world_b.entities.reserve_entities(world_a_max_entities);
Spawn specific entities: spawn or insert operations, refactor spawn internals, world clearing (#2673) This upstreams the code changes used by the new renderer to enable cross-app Entity reuse: * Spawning at specific entities * get_or_spawn: spawns an entity if it doesn't already exist and returns an EntityMut * insert_or_spawn_batch: the batched equivalent to `world.get_or_spawn(entity).insert_bundle(bundle)` * Clearing entities and storages * Allocating Entities with "invalid" archetypes. These entities cannot be queried / are treated as "non existent". They serve as "reserved" entities that won't show up when calling `spawn()`. They must be "specifically spawned at" using apis like `get_or_spawn(entity)`. In combination, these changes enable the "render world" to clear entities / storages each frame and reserve all "app world entities". These can then be spawned during the "render extract step". This refactors "spawn" and "insert" code in a way that I think is a massive improvement to legibility and re-usability. It also yields marginal performance wins by reducing some duplicate lookups (less than a percentage point improvement on insertion benchmarks). There is also some potential for future unsafe reduction (by making BatchSpawner and BatchInserter generic). But for now I want to cut down generic usage to a minimum to encourage smaller binaries and faster compiles. This is currently a draft because it needs more tests (although this code has already had some real-world testing on my custom-shaders branch). I also fixed the benchmarks (which currently don't compile!) / added new ones to illustrate batching wins. After these changes, Bevy ECS is basically ready to accommodate the new renderer. I think the biggest missing piece at this point is "sub apps".
2021-08-25 23:34:02 +00:00
world_b.entities.flush_as_invalid();
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
let e4 = world_b.spawn(A(4)).id();
Spawn specific entities: spawn or insert operations, refactor spawn internals, world clearing (#2673) This upstreams the code changes used by the new renderer to enable cross-app Entity reuse: * Spawning at specific entities * get_or_spawn: spawns an entity if it doesn't already exist and returns an EntityMut * insert_or_spawn_batch: the batched equivalent to `world.get_or_spawn(entity).insert_bundle(bundle)` * Clearing entities and storages * Allocating Entities with "invalid" archetypes. These entities cannot be queried / are treated as "non existent". They serve as "reserved" entities that won't show up when calling `spawn()`. They must be "specifically spawned at" using apis like `get_or_spawn(entity)`. In combination, these changes enable the "render world" to clear entities / storages each frame and reserve all "app world entities". These can then be spawned during the "render extract step". This refactors "spawn" and "insert" code in a way that I think is a massive improvement to legibility and re-usability. It also yields marginal performance wins by reducing some duplicate lookups (less than a percentage point improvement on insertion benchmarks). There is also some potential for future unsafe reduction (by making BatchSpawner and BatchInserter generic). But for now I want to cut down generic usage to a minimum to encourage smaller binaries and faster compiles. This is currently a draft because it needs more tests (although this code has already had some real-world testing on my custom-shaders branch). I also fixed the benchmarks (which currently don't compile!) / added new ones to illustrate batching wins. After these changes, Bevy ECS is basically ready to accommodate the new renderer. I think the biggest missing piece at this point is "sub apps".
2021-08-25 23:34:02 +00:00
assert_eq!(
e4,
Change Entity::generation from u32 to NonZeroU32 for niche optimization (#9907) # Objective - Implements change described in https://github.com/bevyengine/bevy/issues/3022 - Goal is to allow Entity to benefit from niche optimization, especially in the case of Option<Entity> to reduce memory overhead with structures with empty slots ## Discussion - First PR attempt: https://github.com/bevyengine/bevy/pull/3029 - Discord: https://discord.com/channels/691052431525675048/1154573759752183808/1154573764240093224 ## Solution - Change `Entity::generation` from u32 to NonZeroU32 to allow for niche optimization. - The reason for changing generation rather than index is so that the costs are only encountered on Entity free, instead of on Entity alloc - There was some concern with generations being used, due to there being some desire to introduce flags. This was more to do with the original retirement approach, however, in reality even if generations were reduced to 24-bits, we would still have 16 million generations available before wrapping and current ideas indicate that we would be using closer to 4-bits for flags. - Additionally, another concern was the representation of relationships where NonZeroU32 prevents us using the full address space, talking with Joy it seems unlikely to be an issue. The majority of the time these entity references will be low-index entries (ie. `ChildOf`, `Owes`), these will be able to be fast lookups, and the remainder of the range can use slower lookups to map to the address space. - It has the additional benefit of being less visible to most users, since generation is only ever really set through `from_bits` type methods. - `EntityMeta` was changed to match - On free, generation now explicitly wraps: - Originally, generation would panic in debug mode and wrap in release mode due to using regular ops. - The first attempt at this PR changed the behavior to "retire" slots and remove them from use when generations overflowed. This change was controversial, and likely needs a proper RFC/discussion. - Wrapping matches current release behaviour, and should therefore be less controversial. - Wrapping also more easily migrates to the retirement approach, as users likely to exhaust the exorbitant supply of generations will code defensively against aliasing and that defensive code is less likely to break than code assuming that generations don't wrap. - We use some unsafe code here when wrapping generations, to avoid branch on NonZeroU32 construction. It's guaranteed safe due to how we perform wrapping and it results in significantly smaller ASM code. - https://godbolt.org/z/6b6hj8PrM ## Migration - Previous `bevy_scene` serializations have a high likelihood of being broken, as they contain 0th generation entities. ## Current Issues - `Entities::reserve_generations` and `EntityMapper` wrap now, even in debug - although they technically did in release mode already so this probably isn't a huge issue. It just depends if we need to change anything here? --------- Co-authored-by: Natalie Baker <natalie.baker@advancednavigation.com>
2024-01-08 23:03:00 +00:00
Entity::from_raw(3),
Spawn specific entities: spawn or insert operations, refactor spawn internals, world clearing (#2673) This upstreams the code changes used by the new renderer to enable cross-app Entity reuse: * Spawning at specific entities * get_or_spawn: spawns an entity if it doesn't already exist and returns an EntityMut * insert_or_spawn_batch: the batched equivalent to `world.get_or_spawn(entity).insert_bundle(bundle)` * Clearing entities and storages * Allocating Entities with "invalid" archetypes. These entities cannot be queried / are treated as "non existent". They serve as "reserved" entities that won't show up when calling `spawn()`. They must be "specifically spawned at" using apis like `get_or_spawn(entity)`. In combination, these changes enable the "render world" to clear entities / storages each frame and reserve all "app world entities". These can then be spawned during the "render extract step". This refactors "spawn" and "insert" code in a way that I think is a massive improvement to legibility and re-usability. It also yields marginal performance wins by reducing some duplicate lookups (less than a percentage point improvement on insertion benchmarks). There is also some potential for future unsafe reduction (by making BatchSpawner and BatchInserter generic). But for now I want to cut down generic usage to a minimum to encourage smaller binaries and faster compiles. This is currently a draft because it needs more tests (although this code has already had some real-world testing on my custom-shaders branch). I also fixed the benchmarks (which currently don't compile!) / added new ones to illustrate batching wins. After these changes, Bevy ECS is basically ready to accommodate the new renderer. I think the biggest missing piece at this point is "sub apps".
2021-08-25 23:34:02 +00:00
"new entity is created immediately after world_a's max entity"
);
assert!(world_b.get::<A>(e1).is_none());
Spawn specific entities: spawn or insert operations, refactor spawn internals, world clearing (#2673) This upstreams the code changes used by the new renderer to enable cross-app Entity reuse: * Spawning at specific entities * get_or_spawn: spawns an entity if it doesn't already exist and returns an EntityMut * insert_or_spawn_batch: the batched equivalent to `world.get_or_spawn(entity).insert_bundle(bundle)` * Clearing entities and storages * Allocating Entities with "invalid" archetypes. These entities cannot be queried / are treated as "non existent". They serve as "reserved" entities that won't show up when calling `spawn()`. They must be "specifically spawned at" using apis like `get_or_spawn(entity)`. In combination, these changes enable the "render world" to clear entities / storages each frame and reserve all "app world entities". These can then be spawned during the "render extract step". This refactors "spawn" and "insert" code in a way that I think is a massive improvement to legibility and re-usability. It also yields marginal performance wins by reducing some duplicate lookups (less than a percentage point improvement on insertion benchmarks). There is also some potential for future unsafe reduction (by making BatchSpawner and BatchInserter generic). But for now I want to cut down generic usage to a minimum to encourage smaller binaries and faster compiles. This is currently a draft because it needs more tests (although this code has already had some real-world testing on my custom-shaders branch). I also fixed the benchmarks (which currently don't compile!) / added new ones to illustrate batching wins. After these changes, Bevy ECS is basically ready to accommodate the new renderer. I think the biggest missing piece at this point is "sub apps".
2021-08-25 23:34:02 +00:00
assert!(world_b.get_entity(e1).is_none());
assert!(world_b.get::<A>(e2).is_none());
Spawn specific entities: spawn or insert operations, refactor spawn internals, world clearing (#2673) This upstreams the code changes used by the new renderer to enable cross-app Entity reuse: * Spawning at specific entities * get_or_spawn: spawns an entity if it doesn't already exist and returns an EntityMut * insert_or_spawn_batch: the batched equivalent to `world.get_or_spawn(entity).insert_bundle(bundle)` * Clearing entities and storages * Allocating Entities with "invalid" archetypes. These entities cannot be queried / are treated as "non existent". They serve as "reserved" entities that won't show up when calling `spawn()`. They must be "specifically spawned at" using apis like `get_or_spawn(entity)`. In combination, these changes enable the "render world" to clear entities / storages each frame and reserve all "app world entities". These can then be spawned during the "render extract step". This refactors "spawn" and "insert" code in a way that I think is a massive improvement to legibility and re-usability. It also yields marginal performance wins by reducing some duplicate lookups (less than a percentage point improvement on insertion benchmarks). There is also some potential for future unsafe reduction (by making BatchSpawner and BatchInserter generic). But for now I want to cut down generic usage to a minimum to encourage smaller binaries and faster compiles. This is currently a draft because it needs more tests (although this code has already had some real-world testing on my custom-shaders branch). I also fixed the benchmarks (which currently don't compile!) / added new ones to illustrate batching wins. After these changes, Bevy ECS is basically ready to accommodate the new renderer. I think the biggest missing piece at this point is "sub apps".
2021-08-25 23:34:02 +00:00
assert!(world_b.get_entity(e2).is_none());
assert!(world_b.get::<A>(e3).is_none());
Spawn specific entities: spawn or insert operations, refactor spawn internals, world clearing (#2673) This upstreams the code changes used by the new renderer to enable cross-app Entity reuse: * Spawning at specific entities * get_or_spawn: spawns an entity if it doesn't already exist and returns an EntityMut * insert_or_spawn_batch: the batched equivalent to `world.get_or_spawn(entity).insert_bundle(bundle)` * Clearing entities and storages * Allocating Entities with "invalid" archetypes. These entities cannot be queried / are treated as "non existent". They serve as "reserved" entities that won't show up when calling `spawn()`. They must be "specifically spawned at" using apis like `get_or_spawn(entity)`. In combination, these changes enable the "render world" to clear entities / storages each frame and reserve all "app world entities". These can then be spawned during the "render extract step". This refactors "spawn" and "insert" code in a way that I think is a massive improvement to legibility and re-usability. It also yields marginal performance wins by reducing some duplicate lookups (less than a percentage point improvement on insertion benchmarks). There is also some potential for future unsafe reduction (by making BatchSpawner and BatchInserter generic). But for now I want to cut down generic usage to a minimum to encourage smaller binaries and faster compiles. This is currently a draft because it needs more tests (although this code has already had some real-world testing on my custom-shaders branch). I also fixed the benchmarks (which currently don't compile!) / added new ones to illustrate batching wins. After these changes, Bevy ECS is basically ready to accommodate the new renderer. I think the biggest missing piece at this point is "sub apps".
2021-08-25 23:34:02 +00:00
assert!(world_b.get_entity(e3).is_none());
world_b.get_or_spawn(e1).unwrap().insert(B(1));
Spawn specific entities: spawn or insert operations, refactor spawn internals, world clearing (#2673) This upstreams the code changes used by the new renderer to enable cross-app Entity reuse: * Spawning at specific entities * get_or_spawn: spawns an entity if it doesn't already exist and returns an EntityMut * insert_or_spawn_batch: the batched equivalent to `world.get_or_spawn(entity).insert_bundle(bundle)` * Clearing entities and storages * Allocating Entities with "invalid" archetypes. These entities cannot be queried / are treated as "non existent". They serve as "reserved" entities that won't show up when calling `spawn()`. They must be "specifically spawned at" using apis like `get_or_spawn(entity)`. In combination, these changes enable the "render world" to clear entities / storages each frame and reserve all "app world entities". These can then be spawned during the "render extract step". This refactors "spawn" and "insert" code in a way that I think is a massive improvement to legibility and re-usability. It also yields marginal performance wins by reducing some duplicate lookups (less than a percentage point improvement on insertion benchmarks). There is also some potential for future unsafe reduction (by making BatchSpawner and BatchInserter generic). But for now I want to cut down generic usage to a minimum to encourage smaller binaries and faster compiles. This is currently a draft because it needs more tests (although this code has already had some real-world testing on my custom-shaders branch). I also fixed the benchmarks (which currently don't compile!) / added new ones to illustrate batching wins. After these changes, Bevy ECS is basically ready to accommodate the new renderer. I think the biggest missing piece at this point is "sub apps".
2021-08-25 23:34:02 +00:00
assert_eq!(
world_b.get::<B>(e1),
Some(&B(1)),
Spawn specific entities: spawn or insert operations, refactor spawn internals, world clearing (#2673) This upstreams the code changes used by the new renderer to enable cross-app Entity reuse: * Spawning at specific entities * get_or_spawn: spawns an entity if it doesn't already exist and returns an EntityMut * insert_or_spawn_batch: the batched equivalent to `world.get_or_spawn(entity).insert_bundle(bundle)` * Clearing entities and storages * Allocating Entities with "invalid" archetypes. These entities cannot be queried / are treated as "non existent". They serve as "reserved" entities that won't show up when calling `spawn()`. They must be "specifically spawned at" using apis like `get_or_spawn(entity)`. In combination, these changes enable the "render world" to clear entities / storages each frame and reserve all "app world entities". These can then be spawned during the "render extract step". This refactors "spawn" and "insert" code in a way that I think is a massive improvement to legibility and re-usability. It also yields marginal performance wins by reducing some duplicate lookups (less than a percentage point improvement on insertion benchmarks). There is also some potential for future unsafe reduction (by making BatchSpawner and BatchInserter generic). But for now I want to cut down generic usage to a minimum to encourage smaller binaries and faster compiles. This is currently a draft because it needs more tests (although this code has already had some real-world testing on my custom-shaders branch). I also fixed the benchmarks (which currently don't compile!) / added new ones to illustrate batching wins. After these changes, Bevy ECS is basically ready to accommodate the new renderer. I think the biggest missing piece at this point is "sub apps".
2021-08-25 23:34:02 +00:00
"spawning into 'world_a' entities works"
);
world_b.get_or_spawn(e4).unwrap().insert(B(4));
Spawn specific entities: spawn or insert operations, refactor spawn internals, world clearing (#2673) This upstreams the code changes used by the new renderer to enable cross-app Entity reuse: * Spawning at specific entities * get_or_spawn: spawns an entity if it doesn't already exist and returns an EntityMut * insert_or_spawn_batch: the batched equivalent to `world.get_or_spawn(entity).insert_bundle(bundle)` * Clearing entities and storages * Allocating Entities with "invalid" archetypes. These entities cannot be queried / are treated as "non existent". They serve as "reserved" entities that won't show up when calling `spawn()`. They must be "specifically spawned at" using apis like `get_or_spawn(entity)`. In combination, these changes enable the "render world" to clear entities / storages each frame and reserve all "app world entities". These can then be spawned during the "render extract step". This refactors "spawn" and "insert" code in a way that I think is a massive improvement to legibility and re-usability. It also yields marginal performance wins by reducing some duplicate lookups (less than a percentage point improvement on insertion benchmarks). There is also some potential for future unsafe reduction (by making BatchSpawner and BatchInserter generic). But for now I want to cut down generic usage to a minimum to encourage smaller binaries and faster compiles. This is currently a draft because it needs more tests (although this code has already had some real-world testing on my custom-shaders branch). I also fixed the benchmarks (which currently don't compile!) / added new ones to illustrate batching wins. After these changes, Bevy ECS is basically ready to accommodate the new renderer. I think the biggest missing piece at this point is "sub apps".
2021-08-25 23:34:02 +00:00
assert_eq!(
world_b.get::<B>(e4),
Some(&B(4)),
Spawn specific entities: spawn or insert operations, refactor spawn internals, world clearing (#2673) This upstreams the code changes used by the new renderer to enable cross-app Entity reuse: * Spawning at specific entities * get_or_spawn: spawns an entity if it doesn't already exist and returns an EntityMut * insert_or_spawn_batch: the batched equivalent to `world.get_or_spawn(entity).insert_bundle(bundle)` * Clearing entities and storages * Allocating Entities with "invalid" archetypes. These entities cannot be queried / are treated as "non existent". They serve as "reserved" entities that won't show up when calling `spawn()`. They must be "specifically spawned at" using apis like `get_or_spawn(entity)`. In combination, these changes enable the "render world" to clear entities / storages each frame and reserve all "app world entities". These can then be spawned during the "render extract step". This refactors "spawn" and "insert" code in a way that I think is a massive improvement to legibility and re-usability. It also yields marginal performance wins by reducing some duplicate lookups (less than a percentage point improvement on insertion benchmarks). There is also some potential for future unsafe reduction (by making BatchSpawner and BatchInserter generic). But for now I want to cut down generic usage to a minimum to encourage smaller binaries and faster compiles. This is currently a draft because it needs more tests (although this code has already had some real-world testing on my custom-shaders branch). I also fixed the benchmarks (which currently don't compile!) / added new ones to illustrate batching wins. After these changes, Bevy ECS is basically ready to accommodate the new renderer. I think the biggest missing piece at this point is "sub apps".
2021-08-25 23:34:02 +00:00
"spawning into existing `world_b` entities works"
);
assert_eq!(
world_b.get::<A>(e4),
Some(&A(4)),
Spawn specific entities: spawn or insert operations, refactor spawn internals, world clearing (#2673) This upstreams the code changes used by the new renderer to enable cross-app Entity reuse: * Spawning at specific entities * get_or_spawn: spawns an entity if it doesn't already exist and returns an EntityMut * insert_or_spawn_batch: the batched equivalent to `world.get_or_spawn(entity).insert_bundle(bundle)` * Clearing entities and storages * Allocating Entities with "invalid" archetypes. These entities cannot be queried / are treated as "non existent". They serve as "reserved" entities that won't show up when calling `spawn()`. They must be "specifically spawned at" using apis like `get_or_spawn(entity)`. In combination, these changes enable the "render world" to clear entities / storages each frame and reserve all "app world entities". These can then be spawned during the "render extract step". This refactors "spawn" and "insert" code in a way that I think is a massive improvement to legibility and re-usability. It also yields marginal performance wins by reducing some duplicate lookups (less than a percentage point improvement on insertion benchmarks). There is also some potential for future unsafe reduction (by making BatchSpawner and BatchInserter generic). But for now I want to cut down generic usage to a minimum to encourage smaller binaries and faster compiles. This is currently a draft because it needs more tests (although this code has already had some real-world testing on my custom-shaders branch). I also fixed the benchmarks (which currently don't compile!) / added new ones to illustrate batching wins. After these changes, Bevy ECS is basically ready to accommodate the new renderer. I think the biggest missing piece at this point is "sub apps".
2021-08-25 23:34:02 +00:00
"spawning into existing `world_b` entities works"
);
Change Entity::generation from u32 to NonZeroU32 for niche optimization (#9907) # Objective - Implements change described in https://github.com/bevyengine/bevy/issues/3022 - Goal is to allow Entity to benefit from niche optimization, especially in the case of Option<Entity> to reduce memory overhead with structures with empty slots ## Discussion - First PR attempt: https://github.com/bevyengine/bevy/pull/3029 - Discord: https://discord.com/channels/691052431525675048/1154573759752183808/1154573764240093224 ## Solution - Change `Entity::generation` from u32 to NonZeroU32 to allow for niche optimization. - The reason for changing generation rather than index is so that the costs are only encountered on Entity free, instead of on Entity alloc - There was some concern with generations being used, due to there being some desire to introduce flags. This was more to do with the original retirement approach, however, in reality even if generations were reduced to 24-bits, we would still have 16 million generations available before wrapping and current ideas indicate that we would be using closer to 4-bits for flags. - Additionally, another concern was the representation of relationships where NonZeroU32 prevents us using the full address space, talking with Joy it seems unlikely to be an issue. The majority of the time these entity references will be low-index entries (ie. `ChildOf`, `Owes`), these will be able to be fast lookups, and the remainder of the range can use slower lookups to map to the address space. - It has the additional benefit of being less visible to most users, since generation is only ever really set through `from_bits` type methods. - `EntityMeta` was changed to match - On free, generation now explicitly wraps: - Originally, generation would panic in debug mode and wrap in release mode due to using regular ops. - The first attempt at this PR changed the behavior to "retire" slots and remove them from use when generations overflowed. This change was controversial, and likely needs a proper RFC/discussion. - Wrapping matches current release behaviour, and should therefore be less controversial. - Wrapping also more easily migrates to the retirement approach, as users likely to exhaust the exorbitant supply of generations will code defensively against aliasing and that defensive code is less likely to break than code assuming that generations don't wrap. - We use some unsafe code here when wrapping generations, to avoid branch on NonZeroU32 construction. It's guaranteed safe due to how we perform wrapping and it results in significantly smaller ASM code. - https://godbolt.org/z/6b6hj8PrM ## Migration - Previous `bevy_scene` serializations have a high likelihood of being broken, as they contain 0th generation entities. ## Current Issues - `Entities::reserve_generations` and `EntityMapper` wrap now, even in debug - although they technically did in release mode already so this probably isn't a huge issue. It just depends if we need to change anything here? --------- Co-authored-by: Natalie Baker <natalie.baker@advancednavigation.com>
2024-01-08 23:03:00 +00:00
let e4_mismatched_generation = Entity::new(3, 2).unwrap();
Spawn specific entities: spawn or insert operations, refactor spawn internals, world clearing (#2673) This upstreams the code changes used by the new renderer to enable cross-app Entity reuse: * Spawning at specific entities * get_or_spawn: spawns an entity if it doesn't already exist and returns an EntityMut * insert_or_spawn_batch: the batched equivalent to `world.get_or_spawn(entity).insert_bundle(bundle)` * Clearing entities and storages * Allocating Entities with "invalid" archetypes. These entities cannot be queried / are treated as "non existent". They serve as "reserved" entities that won't show up when calling `spawn()`. They must be "specifically spawned at" using apis like `get_or_spawn(entity)`. In combination, these changes enable the "render world" to clear entities / storages each frame and reserve all "app world entities". These can then be spawned during the "render extract step". This refactors "spawn" and "insert" code in a way that I think is a massive improvement to legibility and re-usability. It also yields marginal performance wins by reducing some duplicate lookups (less than a percentage point improvement on insertion benchmarks). There is also some potential for future unsafe reduction (by making BatchSpawner and BatchInserter generic). But for now I want to cut down generic usage to a minimum to encourage smaller binaries and faster compiles. This is currently a draft because it needs more tests (although this code has already had some real-world testing on my custom-shaders branch). I also fixed the benchmarks (which currently don't compile!) / added new ones to illustrate batching wins. After these changes, Bevy ECS is basically ready to accommodate the new renderer. I think the biggest missing piece at this point is "sub apps".
2021-08-25 23:34:02 +00:00
assert!(
world_b.get_or_spawn(e4_mismatched_generation).is_none(),
"attempting to spawn on top of an entity with a mismatched entity generation fails"
);
assert_eq!(
world_b.get::<B>(e4),
Some(&B(4)),
Spawn specific entities: spawn or insert operations, refactor spawn internals, world clearing (#2673) This upstreams the code changes used by the new renderer to enable cross-app Entity reuse: * Spawning at specific entities * get_or_spawn: spawns an entity if it doesn't already exist and returns an EntityMut * insert_or_spawn_batch: the batched equivalent to `world.get_or_spawn(entity).insert_bundle(bundle)` * Clearing entities and storages * Allocating Entities with "invalid" archetypes. These entities cannot be queried / are treated as "non existent". They serve as "reserved" entities that won't show up when calling `spawn()`. They must be "specifically spawned at" using apis like `get_or_spawn(entity)`. In combination, these changes enable the "render world" to clear entities / storages each frame and reserve all "app world entities". These can then be spawned during the "render extract step". This refactors "spawn" and "insert" code in a way that I think is a massive improvement to legibility and re-usability. It also yields marginal performance wins by reducing some duplicate lookups (less than a percentage point improvement on insertion benchmarks). There is also some potential for future unsafe reduction (by making BatchSpawner and BatchInserter generic). But for now I want to cut down generic usage to a minimum to encourage smaller binaries and faster compiles. This is currently a draft because it needs more tests (although this code has already had some real-world testing on my custom-shaders branch). I also fixed the benchmarks (which currently don't compile!) / added new ones to illustrate batching wins. After these changes, Bevy ECS is basically ready to accommodate the new renderer. I think the biggest missing piece at this point is "sub apps".
2021-08-25 23:34:02 +00:00
"failed mismatched spawn doesn't change existing entity"
);
assert_eq!(
world_b.get::<A>(e4),
Some(&A(4)),
Spawn specific entities: spawn or insert operations, refactor spawn internals, world clearing (#2673) This upstreams the code changes used by the new renderer to enable cross-app Entity reuse: * Spawning at specific entities * get_or_spawn: spawns an entity if it doesn't already exist and returns an EntityMut * insert_or_spawn_batch: the batched equivalent to `world.get_or_spawn(entity).insert_bundle(bundle)` * Clearing entities and storages * Allocating Entities with "invalid" archetypes. These entities cannot be queried / are treated as "non existent". They serve as "reserved" entities that won't show up when calling `spawn()`. They must be "specifically spawned at" using apis like `get_or_spawn(entity)`. In combination, these changes enable the "render world" to clear entities / storages each frame and reserve all "app world entities". These can then be spawned during the "render extract step". This refactors "spawn" and "insert" code in a way that I think is a massive improvement to legibility and re-usability. It also yields marginal performance wins by reducing some duplicate lookups (less than a percentage point improvement on insertion benchmarks). There is also some potential for future unsafe reduction (by making BatchSpawner and BatchInserter generic). But for now I want to cut down generic usage to a minimum to encourage smaller binaries and faster compiles. This is currently a draft because it needs more tests (although this code has already had some real-world testing on my custom-shaders branch). I also fixed the benchmarks (which currently don't compile!) / added new ones to illustrate batching wins. After these changes, Bevy ECS is basically ready to accommodate the new renderer. I think the biggest missing piece at this point is "sub apps".
2021-08-25 23:34:02 +00:00
"failed mismatched spawn doesn't change existing entity"
);
Change Entity::generation from u32 to NonZeroU32 for niche optimization (#9907) # Objective - Implements change described in https://github.com/bevyengine/bevy/issues/3022 - Goal is to allow Entity to benefit from niche optimization, especially in the case of Option<Entity> to reduce memory overhead with structures with empty slots ## Discussion - First PR attempt: https://github.com/bevyengine/bevy/pull/3029 - Discord: https://discord.com/channels/691052431525675048/1154573759752183808/1154573764240093224 ## Solution - Change `Entity::generation` from u32 to NonZeroU32 to allow for niche optimization. - The reason for changing generation rather than index is so that the costs are only encountered on Entity free, instead of on Entity alloc - There was some concern with generations being used, due to there being some desire to introduce flags. This was more to do with the original retirement approach, however, in reality even if generations were reduced to 24-bits, we would still have 16 million generations available before wrapping and current ideas indicate that we would be using closer to 4-bits for flags. - Additionally, another concern was the representation of relationships where NonZeroU32 prevents us using the full address space, talking with Joy it seems unlikely to be an issue. The majority of the time these entity references will be low-index entries (ie. `ChildOf`, `Owes`), these will be able to be fast lookups, and the remainder of the range can use slower lookups to map to the address space. - It has the additional benefit of being less visible to most users, since generation is only ever really set through `from_bits` type methods. - `EntityMeta` was changed to match - On free, generation now explicitly wraps: - Originally, generation would panic in debug mode and wrap in release mode due to using regular ops. - The first attempt at this PR changed the behavior to "retire" slots and remove them from use when generations overflowed. This change was controversial, and likely needs a proper RFC/discussion. - Wrapping matches current release behaviour, and should therefore be less controversial. - Wrapping also more easily migrates to the retirement approach, as users likely to exhaust the exorbitant supply of generations will code defensively against aliasing and that defensive code is less likely to break than code assuming that generations don't wrap. - We use some unsafe code here when wrapping generations, to avoid branch on NonZeroU32 construction. It's guaranteed safe due to how we perform wrapping and it results in significantly smaller ASM code. - https://godbolt.org/z/6b6hj8PrM ## Migration - Previous `bevy_scene` serializations have a high likelihood of being broken, as they contain 0th generation entities. ## Current Issues - `Entities::reserve_generations` and `EntityMapper` wrap now, even in debug - although they technically did in release mode already so this probably isn't a huge issue. It just depends if we need to change anything here? --------- Co-authored-by: Natalie Baker <natalie.baker@advancednavigation.com>
2024-01-08 23:03:00 +00:00
let high_non_existent_entity = Entity::from_raw(6);
Spawn specific entities: spawn or insert operations, refactor spawn internals, world clearing (#2673) This upstreams the code changes used by the new renderer to enable cross-app Entity reuse: * Spawning at specific entities * get_or_spawn: spawns an entity if it doesn't already exist and returns an EntityMut * insert_or_spawn_batch: the batched equivalent to `world.get_or_spawn(entity).insert_bundle(bundle)` * Clearing entities and storages * Allocating Entities with "invalid" archetypes. These entities cannot be queried / are treated as "non existent". They serve as "reserved" entities that won't show up when calling `spawn()`. They must be "specifically spawned at" using apis like `get_or_spawn(entity)`. In combination, these changes enable the "render world" to clear entities / storages each frame and reserve all "app world entities". These can then be spawned during the "render extract step". This refactors "spawn" and "insert" code in a way that I think is a massive improvement to legibility and re-usability. It also yields marginal performance wins by reducing some duplicate lookups (less than a percentage point improvement on insertion benchmarks). There is also some potential for future unsafe reduction (by making BatchSpawner and BatchInserter generic). But for now I want to cut down generic usage to a minimum to encourage smaller binaries and faster compiles. This is currently a draft because it needs more tests (although this code has already had some real-world testing on my custom-shaders branch). I also fixed the benchmarks (which currently don't compile!) / added new ones to illustrate batching wins. After these changes, Bevy ECS is basically ready to accommodate the new renderer. I think the biggest missing piece at this point is "sub apps".
2021-08-25 23:34:02 +00:00
world_b
.get_or_spawn(high_non_existent_entity)
.unwrap()
.insert(B(10));
Spawn specific entities: spawn or insert operations, refactor spawn internals, world clearing (#2673) This upstreams the code changes used by the new renderer to enable cross-app Entity reuse: * Spawning at specific entities * get_or_spawn: spawns an entity if it doesn't already exist and returns an EntityMut * insert_or_spawn_batch: the batched equivalent to `world.get_or_spawn(entity).insert_bundle(bundle)` * Clearing entities and storages * Allocating Entities with "invalid" archetypes. These entities cannot be queried / are treated as "non existent". They serve as "reserved" entities that won't show up when calling `spawn()`. They must be "specifically spawned at" using apis like `get_or_spawn(entity)`. In combination, these changes enable the "render world" to clear entities / storages each frame and reserve all "app world entities". These can then be spawned during the "render extract step". This refactors "spawn" and "insert" code in a way that I think is a massive improvement to legibility and re-usability. It also yields marginal performance wins by reducing some duplicate lookups (less than a percentage point improvement on insertion benchmarks). There is also some potential for future unsafe reduction (by making BatchSpawner and BatchInserter generic). But for now I want to cut down generic usage to a minimum to encourage smaller binaries and faster compiles. This is currently a draft because it needs more tests (although this code has already had some real-world testing on my custom-shaders branch). I also fixed the benchmarks (which currently don't compile!) / added new ones to illustrate batching wins. After these changes, Bevy ECS is basically ready to accommodate the new renderer. I think the biggest missing piece at this point is "sub apps".
2021-08-25 23:34:02 +00:00
assert_eq!(
world_b.get::<B>(high_non_existent_entity),
Some(&B(10)),
"inserting into newly allocated high / non-continuous entity id works"
Spawn specific entities: spawn or insert operations, refactor spawn internals, world clearing (#2673) This upstreams the code changes used by the new renderer to enable cross-app Entity reuse: * Spawning at specific entities * get_or_spawn: spawns an entity if it doesn't already exist and returns an EntityMut * insert_or_spawn_batch: the batched equivalent to `world.get_or_spawn(entity).insert_bundle(bundle)` * Clearing entities and storages * Allocating Entities with "invalid" archetypes. These entities cannot be queried / are treated as "non existent". They serve as "reserved" entities that won't show up when calling `spawn()`. They must be "specifically spawned at" using apis like `get_or_spawn(entity)`. In combination, these changes enable the "render world" to clear entities / storages each frame and reserve all "app world entities". These can then be spawned during the "render extract step". This refactors "spawn" and "insert" code in a way that I think is a massive improvement to legibility and re-usability. It also yields marginal performance wins by reducing some duplicate lookups (less than a percentage point improvement on insertion benchmarks). There is also some potential for future unsafe reduction (by making BatchSpawner and BatchInserter generic). But for now I want to cut down generic usage to a minimum to encourage smaller binaries and faster compiles. This is currently a draft because it needs more tests (although this code has already had some real-world testing on my custom-shaders branch). I also fixed the benchmarks (which currently don't compile!) / added new ones to illustrate batching wins. After these changes, Bevy ECS is basically ready to accommodate the new renderer. I think the biggest missing piece at this point is "sub apps".
2021-08-25 23:34:02 +00:00
);
Change Entity::generation from u32 to NonZeroU32 for niche optimization (#9907) # Objective - Implements change described in https://github.com/bevyengine/bevy/issues/3022 - Goal is to allow Entity to benefit from niche optimization, especially in the case of Option<Entity> to reduce memory overhead with structures with empty slots ## Discussion - First PR attempt: https://github.com/bevyengine/bevy/pull/3029 - Discord: https://discord.com/channels/691052431525675048/1154573759752183808/1154573764240093224 ## Solution - Change `Entity::generation` from u32 to NonZeroU32 to allow for niche optimization. - The reason for changing generation rather than index is so that the costs are only encountered on Entity free, instead of on Entity alloc - There was some concern with generations being used, due to there being some desire to introduce flags. This was more to do with the original retirement approach, however, in reality even if generations were reduced to 24-bits, we would still have 16 million generations available before wrapping and current ideas indicate that we would be using closer to 4-bits for flags. - Additionally, another concern was the representation of relationships where NonZeroU32 prevents us using the full address space, talking with Joy it seems unlikely to be an issue. The majority of the time these entity references will be low-index entries (ie. `ChildOf`, `Owes`), these will be able to be fast lookups, and the remainder of the range can use slower lookups to map to the address space. - It has the additional benefit of being less visible to most users, since generation is only ever really set through `from_bits` type methods. - `EntityMeta` was changed to match - On free, generation now explicitly wraps: - Originally, generation would panic in debug mode and wrap in release mode due to using regular ops. - The first attempt at this PR changed the behavior to "retire" slots and remove them from use when generations overflowed. This change was controversial, and likely needs a proper RFC/discussion. - Wrapping matches current release behaviour, and should therefore be less controversial. - Wrapping also more easily migrates to the retirement approach, as users likely to exhaust the exorbitant supply of generations will code defensively against aliasing and that defensive code is less likely to break than code assuming that generations don't wrap. - We use some unsafe code here when wrapping generations, to avoid branch on NonZeroU32 construction. It's guaranteed safe due to how we perform wrapping and it results in significantly smaller ASM code. - https://godbolt.org/z/6b6hj8PrM ## Migration - Previous `bevy_scene` serializations have a high likelihood of being broken, as they contain 0th generation entities. ## Current Issues - `Entities::reserve_generations` and `EntityMapper` wrap now, even in debug - although they technically did in release mode already so this probably isn't a huge issue. It just depends if we need to change anything here? --------- Co-authored-by: Natalie Baker <natalie.baker@advancednavigation.com>
2024-01-08 23:03:00 +00:00
let high_non_existent_but_reserved_entity = Entity::from_raw(5);
Spawn specific entities: spawn or insert operations, refactor spawn internals, world clearing (#2673) This upstreams the code changes used by the new renderer to enable cross-app Entity reuse: * Spawning at specific entities * get_or_spawn: spawns an entity if it doesn't already exist and returns an EntityMut * insert_or_spawn_batch: the batched equivalent to `world.get_or_spawn(entity).insert_bundle(bundle)` * Clearing entities and storages * Allocating Entities with "invalid" archetypes. These entities cannot be queried / are treated as "non existent". They serve as "reserved" entities that won't show up when calling `spawn()`. They must be "specifically spawned at" using apis like `get_or_spawn(entity)`. In combination, these changes enable the "render world" to clear entities / storages each frame and reserve all "app world entities". These can then be spawned during the "render extract step". This refactors "spawn" and "insert" code in a way that I think is a massive improvement to legibility and re-usability. It also yields marginal performance wins by reducing some duplicate lookups (less than a percentage point improvement on insertion benchmarks). There is also some potential for future unsafe reduction (by making BatchSpawner and BatchInserter generic). But for now I want to cut down generic usage to a minimum to encourage smaller binaries and faster compiles. This is currently a draft because it needs more tests (although this code has already had some real-world testing on my custom-shaders branch). I also fixed the benchmarks (which currently don't compile!) / added new ones to illustrate batching wins. After these changes, Bevy ECS is basically ready to accommodate the new renderer. I think the biggest missing piece at this point is "sub apps".
2021-08-25 23:34:02 +00:00
assert!(
world_b.get_entity(high_non_existent_but_reserved_entity).is_none(),
"entities between high-newly allocated entity and continuous block of existing entities don't exist"
);
let reserved_entities = vec![
world_b.entities().reserve_entity(),
world_b.entities().reserve_entity(),
world_b.entities().reserve_entity(),
world_b.entities().reserve_entity(),
];
assert_eq!(
reserved_entities,
vec![
Change Entity::generation from u32 to NonZeroU32 for niche optimization (#9907) # Objective - Implements change described in https://github.com/bevyengine/bevy/issues/3022 - Goal is to allow Entity to benefit from niche optimization, especially in the case of Option<Entity> to reduce memory overhead with structures with empty slots ## Discussion - First PR attempt: https://github.com/bevyengine/bevy/pull/3029 - Discord: https://discord.com/channels/691052431525675048/1154573759752183808/1154573764240093224 ## Solution - Change `Entity::generation` from u32 to NonZeroU32 to allow for niche optimization. - The reason for changing generation rather than index is so that the costs are only encountered on Entity free, instead of on Entity alloc - There was some concern with generations being used, due to there being some desire to introduce flags. This was more to do with the original retirement approach, however, in reality even if generations were reduced to 24-bits, we would still have 16 million generations available before wrapping and current ideas indicate that we would be using closer to 4-bits for flags. - Additionally, another concern was the representation of relationships where NonZeroU32 prevents us using the full address space, talking with Joy it seems unlikely to be an issue. The majority of the time these entity references will be low-index entries (ie. `ChildOf`, `Owes`), these will be able to be fast lookups, and the remainder of the range can use slower lookups to map to the address space. - It has the additional benefit of being less visible to most users, since generation is only ever really set through `from_bits` type methods. - `EntityMeta` was changed to match - On free, generation now explicitly wraps: - Originally, generation would panic in debug mode and wrap in release mode due to using regular ops. - The first attempt at this PR changed the behavior to "retire" slots and remove them from use when generations overflowed. This change was controversial, and likely needs a proper RFC/discussion. - Wrapping matches current release behaviour, and should therefore be less controversial. - Wrapping also more easily migrates to the retirement approach, as users likely to exhaust the exorbitant supply of generations will code defensively against aliasing and that defensive code is less likely to break than code assuming that generations don't wrap. - We use some unsafe code here when wrapping generations, to avoid branch on NonZeroU32 construction. It's guaranteed safe due to how we perform wrapping and it results in significantly smaller ASM code. - https://godbolt.org/z/6b6hj8PrM ## Migration - Previous `bevy_scene` serializations have a high likelihood of being broken, as they contain 0th generation entities. ## Current Issues - `Entities::reserve_generations` and `EntityMapper` wrap now, even in debug - although they technically did in release mode already so this probably isn't a huge issue. It just depends if we need to change anything here? --------- Co-authored-by: Natalie Baker <natalie.baker@advancednavigation.com>
2024-01-08 23:03:00 +00:00
Entity::from_raw(5),
Entity::from_raw(4),
Entity::from_raw(7),
Entity::from_raw(8),
Spawn specific entities: spawn or insert operations, refactor spawn internals, world clearing (#2673) This upstreams the code changes used by the new renderer to enable cross-app Entity reuse: * Spawning at specific entities * get_or_spawn: spawns an entity if it doesn't already exist and returns an EntityMut * insert_or_spawn_batch: the batched equivalent to `world.get_or_spawn(entity).insert_bundle(bundle)` * Clearing entities and storages * Allocating Entities with "invalid" archetypes. These entities cannot be queried / are treated as "non existent". They serve as "reserved" entities that won't show up when calling `spawn()`. They must be "specifically spawned at" using apis like `get_or_spawn(entity)`. In combination, these changes enable the "render world" to clear entities / storages each frame and reserve all "app world entities". These can then be spawned during the "render extract step". This refactors "spawn" and "insert" code in a way that I think is a massive improvement to legibility and re-usability. It also yields marginal performance wins by reducing some duplicate lookups (less than a percentage point improvement on insertion benchmarks). There is also some potential for future unsafe reduction (by making BatchSpawner and BatchInserter generic). But for now I want to cut down generic usage to a minimum to encourage smaller binaries and faster compiles. This is currently a draft because it needs more tests (although this code has already had some real-world testing on my custom-shaders branch). I also fixed the benchmarks (which currently don't compile!) / added new ones to illustrate batching wins. After these changes, Bevy ECS is basically ready to accommodate the new renderer. I think the biggest missing piece at this point is "sub apps".
2021-08-25 23:34:02 +00:00
],
"space between original entities and high entities is used for new entity ids"
);
}
#[test]
fn insert_or_spawn_batch() {
let mut world = World::default();
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
let e0 = world.spawn(A(0)).id();
let e1 = Entity::from_raw(1);
Spawn specific entities: spawn or insert operations, refactor spawn internals, world clearing (#2673) This upstreams the code changes used by the new renderer to enable cross-app Entity reuse: * Spawning at specific entities * get_or_spawn: spawns an entity if it doesn't already exist and returns an EntityMut * insert_or_spawn_batch: the batched equivalent to `world.get_or_spawn(entity).insert_bundle(bundle)` * Clearing entities and storages * Allocating Entities with "invalid" archetypes. These entities cannot be queried / are treated as "non existent". They serve as "reserved" entities that won't show up when calling `spawn()`. They must be "specifically spawned at" using apis like `get_or_spawn(entity)`. In combination, these changes enable the "render world" to clear entities / storages each frame and reserve all "app world entities". These can then be spawned during the "render extract step". This refactors "spawn" and "insert" code in a way that I think is a massive improvement to legibility and re-usability. It also yields marginal performance wins by reducing some duplicate lookups (less than a percentage point improvement on insertion benchmarks). There is also some potential for future unsafe reduction (by making BatchSpawner and BatchInserter generic). But for now I want to cut down generic usage to a minimum to encourage smaller binaries and faster compiles. This is currently a draft because it needs more tests (although this code has already had some real-world testing on my custom-shaders branch). I also fixed the benchmarks (which currently don't compile!) / added new ones to illustrate batching wins. After these changes, Bevy ECS is basically ready to accommodate the new renderer. I think the biggest missing piece at this point is "sub apps".
2021-08-25 23:34:02 +00:00
let values = vec![(e0, (B(0), C)), (e1, (B(1), C))];
world.insert_or_spawn_batch(values).unwrap();
assert_eq!(
world.get::<A>(e0),
Some(&A(0)),
"existing component was preserved"
);
assert_eq!(
world.get::<B>(e0),
Some(&B(0)),
"pre-existing entity received correct B component"
);
assert_eq!(
world.get::<B>(e1),
Some(&B(1)),
"new entity was spawned and received correct B component"
);
assert_eq!(
world.get::<C>(e0),
Some(&C),
"pre-existing entity received C component"
);
assert_eq!(
world.get::<C>(e1),
Some(&C),
"new entity was spawned and received C component"
);
}
#[test]
fn insert_or_spawn_batch_invalid() {
let mut world = World::default();
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
let e0 = world.spawn(A(0)).id();
let e1 = Entity::from_raw(1);
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
let e2 = world.spawn_empty().id();
Change Entity::generation from u32 to NonZeroU32 for niche optimization (#9907) # Objective - Implements change described in https://github.com/bevyengine/bevy/issues/3022 - Goal is to allow Entity to benefit from niche optimization, especially in the case of Option<Entity> to reduce memory overhead with structures with empty slots ## Discussion - First PR attempt: https://github.com/bevyengine/bevy/pull/3029 - Discord: https://discord.com/channels/691052431525675048/1154573759752183808/1154573764240093224 ## Solution - Change `Entity::generation` from u32 to NonZeroU32 to allow for niche optimization. - The reason for changing generation rather than index is so that the costs are only encountered on Entity free, instead of on Entity alloc - There was some concern with generations being used, due to there being some desire to introduce flags. This was more to do with the original retirement approach, however, in reality even if generations were reduced to 24-bits, we would still have 16 million generations available before wrapping and current ideas indicate that we would be using closer to 4-bits for flags. - Additionally, another concern was the representation of relationships where NonZeroU32 prevents us using the full address space, talking with Joy it seems unlikely to be an issue. The majority of the time these entity references will be low-index entries (ie. `ChildOf`, `Owes`), these will be able to be fast lookups, and the remainder of the range can use slower lookups to map to the address space. - It has the additional benefit of being less visible to most users, since generation is only ever really set through `from_bits` type methods. - `EntityMeta` was changed to match - On free, generation now explicitly wraps: - Originally, generation would panic in debug mode and wrap in release mode due to using regular ops. - The first attempt at this PR changed the behavior to "retire" slots and remove them from use when generations overflowed. This change was controversial, and likely needs a proper RFC/discussion. - Wrapping matches current release behaviour, and should therefore be less controversial. - Wrapping also more easily migrates to the retirement approach, as users likely to exhaust the exorbitant supply of generations will code defensively against aliasing and that defensive code is less likely to break than code assuming that generations don't wrap. - We use some unsafe code here when wrapping generations, to avoid branch on NonZeroU32 construction. It's guaranteed safe due to how we perform wrapping and it results in significantly smaller ASM code. - https://godbolt.org/z/6b6hj8PrM ## Migration - Previous `bevy_scene` serializations have a high likelihood of being broken, as they contain 0th generation entities. ## Current Issues - `Entities::reserve_generations` and `EntityMapper` wrap now, even in debug - although they technically did in release mode already so this probably isn't a huge issue. It just depends if we need to change anything here? --------- Co-authored-by: Natalie Baker <natalie.baker@advancednavigation.com>
2024-01-08 23:03:00 +00:00
let invalid_e2 = Entity::new(e2.index(), 2).unwrap();
Spawn specific entities: spawn or insert operations, refactor spawn internals, world clearing (#2673) This upstreams the code changes used by the new renderer to enable cross-app Entity reuse: * Spawning at specific entities * get_or_spawn: spawns an entity if it doesn't already exist and returns an EntityMut * insert_or_spawn_batch: the batched equivalent to `world.get_or_spawn(entity).insert_bundle(bundle)` * Clearing entities and storages * Allocating Entities with "invalid" archetypes. These entities cannot be queried / are treated as "non existent". They serve as "reserved" entities that won't show up when calling `spawn()`. They must be "specifically spawned at" using apis like `get_or_spawn(entity)`. In combination, these changes enable the "render world" to clear entities / storages each frame and reserve all "app world entities". These can then be spawned during the "render extract step". This refactors "spawn" and "insert" code in a way that I think is a massive improvement to legibility and re-usability. It also yields marginal performance wins by reducing some duplicate lookups (less than a percentage point improvement on insertion benchmarks). There is also some potential for future unsafe reduction (by making BatchSpawner and BatchInserter generic). But for now I want to cut down generic usage to a minimum to encourage smaller binaries and faster compiles. This is currently a draft because it needs more tests (although this code has already had some real-world testing on my custom-shaders branch). I also fixed the benchmarks (which currently don't compile!) / added new ones to illustrate batching wins. After these changes, Bevy ECS is basically ready to accommodate the new renderer. I think the biggest missing piece at this point is "sub apps".
2021-08-25 23:34:02 +00:00
let values = vec![(e0, (B(0), C)), (e1, (B(1), C)), (invalid_e2, (B(2), C))];
let result = world.insert_or_spawn_batch(values);
assert_eq!(
result,
Err(vec![invalid_e2]),
"e2 failed to be spawned or inserted into"
);
assert_eq!(
world.get::<A>(e0),
Some(&A(0)),
"existing component was preserved"
);
assert_eq!(
world.get::<B>(e0),
Some(&B(0)),
"pre-existing entity received correct B component"
);
assert_eq!(
world.get::<B>(e1),
Some(&B(1)),
"new entity was spawned and received correct B component"
);
assert_eq!(
world.get::<C>(e0),
Some(&C),
"pre-existing entity received C component"
);
assert_eq!(
world.get::<C>(e1),
Some(&C),
"new entity was spawned and received C component"
);
}
#[derive(Component)]
struct ComponentA(u32);
#[derive(Component)]
struct ComponentB(u32);
#[derive(Bundle)]
struct Simple(ComponentA);
#[derive(Bundle)]
struct Tuple(Simple, ComponentB);
#[derive(Bundle)]
struct Record {
field0: Simple,
field1: ComponentB,
}
Bevy ECS V2 (#1525) # Bevy ECS V2 This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details: * Complete World rewrite * Multiple component storage types: * Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes) * Sparse Sets: fast add/remove, slower iteration * Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now) * Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364) * Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work) * Archetypes are now "just metadata", component storage is separate * Archetype Graph (for faster archetype changes) * Component Metadata * Configure component storage type * Retrieve information about component size/type/name/layout/send-ness/etc * Components are uniquely identified by a densely packed ComponentId * TypeIds are now totally optional (which should make implementing scripting easier) * Super fast "for_each" query iterators * Merged Resources into World. Resources are now just a special type of component * EntityRef/EntityMut builder apis (more efficient and more ergonomic) * Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere * Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime) * With/Without are still taken into account for conflicts, so this should still be comfy to use * Much simpler `IntoSystem` impl * Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId) * Safety Improvements * Entity reservation uses a normal world reference instead of unsafe transmute * QuerySets no longer transmute lifetimes * Made traits "unsafe" where relevant * More thorough safety docs * WorldCell * Exposes safe mutable access to multiple resources at a time in a World * Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)` * Simpler Bundle implementation * Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection" * Removed `Mut<T>` query impl. it is better to only support one way: `&mut T` * Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default * Components now have is_send property (currently only resources support non-send) * More granular module organization * New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()` * `world.resource_scope()` for mutable access to resources and world at the same time * WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it * Significantly slimmed down SystemState in favor of individual SystemParam state * System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference) Fixes #1320 ## `World` Rewrite This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own! (the only shared code between the projects is the entity id allocator, which is already basically ideal) A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details. ## Component Storage (The Problem) Two ECS storage paradigms have gained a lot of traction over the years: * **Archetypal ECS**: * Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity. * Each "archetype" has its own table. Adding/removing an entity's component changes the archetype. * Enables super-fast Query iteration due to its cache-friendly data layout * Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table" * **Sparse Set ECS**: * Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids) * Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array. * Adding/removing components is a cheap, constant time operation Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate. Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because: 1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform. 2. users need to take manual action to optimize Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance. ## Hybrid Component Storage (The Solution) In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed): * **Tables** (aka "archetypal" storage) * The default storage. If you don't configure anything, this is what you get * Fast iteration by default * Slower add/remove operations * **Sparse Sets** * Opt-in * Slower iteration * Faster add/remove operations These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set": ```rust world.register_component( ComponentDescriptor::new::<MyComponent>(StorageType::SparseSet) ).unwrap(); ``` ## Archetypes Archetypes are now "just metadata" ... they no longer store components directly. They do store: * The `ComponentId`s of each of the Archetype's components (and that component's storage type) * Archetypes are uniquely defined by their component layouts * For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype. * The `TableId` associated with the archetype * For now each archetype has exactly one table (which can have no components), * There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it: * Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components. * This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later) * A list of entities that are in the archetype and the row id of the table they are in * ArchetypeComponentIds * unique densely packed identifiers for (ArchetypeId, ComponentId) pairs * used by the schedule executor for cheap system access control * "Archetype Graph Edges" (see the next section) ## The "Archetype Graph" Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage. The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes. Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph. As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations. ## Stateful Queries World queries are now stateful. This allows us to: 1. Cache archetype (and table) matches * This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs). 2. Cache Fetch and Filter state * The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed 3. Incrementally build up state * When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes) As a result, the direct `World` query api now looks like this: ```rust let mut query = world.query::<(&A, &mut B)>(); for (a, mut b) in query.iter_mut(&mut world) { } ``` Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world). However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam. ## Stateful SystemParams Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources). SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now. Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params). (credit goes to @DJMcNab for the initial idea and draft pr here #1364) ## Configurable SystemParams @DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters: ```rust fn foo(value: Local<usize>) { } app.add_system(foo.system().config(|c| c.0 = Some(10))); ``` ## Uber Fast "for_each" Query Iterators Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration. ```rust fn system(query: Query<(&A, &mut B)>) { // you now have the option to do this for a speed boost query.for_each_mut(|(a, mut b)| { }); // however normal iterators are still available for (a, mut b) in query.iter_mut() { } } ``` I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`. We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr). ## Component Metadata `World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type. ## Significantly Cheaper `Access<T>` We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed. This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s. ## Merged Resources into World Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity). Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state. I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally). This pr merges Resources into World: ```rust world.insert_resource(1); world.insert_resource(2.0); let a = world.get_resource::<i32>().unwrap(); let mut b = world.get_resource_mut::<f64>().unwrap(); *b = 3.0; ``` Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier. _But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably! ## WorldCell WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access: ```rust let world_cell = world.cell(); let a = world_cell.get_resource_mut::<i32>().unwrap(); let b = world_cell.get_resource_mut::<f64>().unwrap(); ``` This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped. World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation. WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer. The api is currently limited to resource access, but it can and should be extended to queries / entity component access. ## Resource Scopes WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal! Instead developers can use a "resource scope" ```rust world.resource_scope(|world: &mut World, a: &mut A| { }) ``` This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation. If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty. ## Query Conflicts Use ComponentId Instead of ArchetypeComponentId For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters: ```rust // these queries will never conflict due to their filters fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) { } ``` But it also has a significant downside: ```rust // these queries will not conflict _until_ an entity with A, B, and C is spawned fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) { } ``` The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing. In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace. To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict. ## EntityRef / EntityMut World entity operations on `main` require that the user passes in an `entity` id to each operation: ```rust let entity = world.spawn((A, )); // create a new entity with A world.get::<A>(entity); world.insert(entity, (B, C)); world.insert_one(entity, D); ``` This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required). These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity: ```rust // spawn now takes no inputs and returns an EntityMut let entity = world.spawn() .insert(A) // insert a single component into the entity .insert_bundle((B, C)) // insert a bundle of components into the entity .id() // id returns the Entity id // Returns EntityMut (or panics if the entity does not exist) world.entity_mut(entity) .insert(D) .insert_bundle(SomeBundle::default()); { // returns EntityRef (or panics if the entity does not exist) let d = world.entity(entity) .get::<D>() // gets the D component .unwrap(); // world.get still exists for ergonomics let d = world.get::<D>(entity).unwrap(); } // These variants return Options if you want to check existence instead of panicing world.get_entity_mut(entity) .unwrap() .insert(E); if let Some(entity_ref) = world.get_entity(entity) { let d = entity_ref.get::<D>().unwrap(); } ``` This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change. ## Safety Improvements * Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute * QuerySets no longer transmutes lifetimes * Made traits "unsafe" when implementing a trait incorrectly could cause unsafety * More thorough safety docs ## RemovedComponents SystemParam The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState: ```rust fn system(removed: RemovedComponents<T>) { for entity in removed.iter() { } } ``` ## Simpler Bundle implementation Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used. ## Unified WorldQuery and QueryFilter types (don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change) WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful). QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool. This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit. ## More Granular Modules World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here). ## Remaining Draft Work (to be done in this pr) * ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~ * ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~ * ~~batch_iter / par_iter (currently stubbed out)~~ * ~~ChangedRes~~ * ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~. * ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~ * ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~ * ~~Nested Bundles (if i find time)~~ ## Potential Future Work * Expand WorldCell to support queries. * Consider not allocating in the empty archetype on `world.spawn()` * ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op * this actually regressed performance last time i tried it, but in theory it should be faster * Optimize SparseSet::insert (see `PERF` comment on insert) * Replace SparseArray `Option<T>` with T::MAX to cut down on branching * would enable cheaper get_unchecked() operations * upstream fixedbitset optimizations * fixedbitset could be allocation free for small block counts (store blocks in a SmallVec) * fixedbitset could have a const constructor * Consider implementing Tags (archetype-specific by-value data that affects archetype identity) * ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different. * this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage. * Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation * all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints) * but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code) * Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell * this is basically just "systems" so maybe it's not worth it * Add more world ops * `world.clear()` * `world.reserve<T: Bundle>(count: usize)` * Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :) * Adapt Commands apis for consistency with new World apis ## Benchmarks key: * `bevy_old`: bevy `main` branch * `bevy`: this branch * `_foreach`: uses an optimized for_each iterator * ` _sparse`: uses sparse set storage (if unspecified assume table storage) * `_system`: runs inside a system (if unspecified assume test happens via direct world ops) ### Simple Insert (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245573-9c3ce100-7795-11eb-9003-bfd41cd5c51f.png) ### Simpler Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245795-ffc70e80-7795-11eb-92fb-3ffad09aabf7.png) ### Fragment Iter (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109245849-0fdeee00-7796-11eb-8d25-eb6b7a682c48.png) ### Sparse Fragmented Iter Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes ![image](https://user-images.githubusercontent.com/2694663/109245916-2b49f900-7796-11eb-9a8f-ed89c203f940.png) ### Schedule (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246428-1fab0200-7797-11eb-8841-1b2161e90fa4.png) ### Add Remove Component (from ecs_bench_suite) ![image](https://user-images.githubusercontent.com/2694663/109246492-39e4e000-7797-11eb-8985-2706bd0495ab.png) ### Add Remove Component Big Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed ![image](https://user-images.githubusercontent.com/2694663/109246517-449f7500-7797-11eb-835e-28b6790daeaa.png) ### Get Component Looks up a single component value a large number of times ![image](https://user-images.githubusercontent.com/2694663/109246129-87ad1880-7796-11eb-9fcb-c38012aa7c70.png)
2021-03-05 07:54:35 +00:00
}