# Objective
- Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene
- Fixes#3605
- Based on top of #4079
This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR.
## Solution
- Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability
- Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time.
- Appropriate shader defs in the shader code to handle the two cases
## Context on some decisions / open questions
- I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately.
- Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants?
Co-authored-by: Carter Anderson <mcanders1@gmail.com>
* Refactor assign_lights_to_clusters to always clear + update clusters, even if the screen size isn't available yet / is zero. This fixes#4167. We still avoid the "expensive" per-light work when the screen size isn't available yet. I also consolidated some logic to eliminate some redundancies.
* Removed _a ton_ of (potentially very large) per-frame reallocations
* Removed `Res<VisiblePointLights>` (a vec) in favor of `Res<GlobalVisiblePointLights>` (a hashmap). We were allocating a new hashmap every frame, the collecting it into a vec every frame, then in another system _re-generating the hashmap_. It is always used like a hashmap, might as well embrace that. We now reuse the same hashmap every frame and dont use any intermediate collections.
* We were re-allocating Clusters aabb and light vectors every frame by re-constructing Clusters every frame. We now re-use the existing collections.
* Reuse per-camera VisiblePointLight vecs when possible instead of allocating them every frame. We now only insert VisiblePointLights if the component doesn't exist yet.
# Objective
- Reduce time spent in the `check_visibility` system
## Solution
- Use `Vec3A` for all bounding volume types to leverage SIMD optimisations and to avoid repeated runtime conversions from `Vec3` to `Vec3A`
- Inline all bounding volume intersection methods
- Add on-the-fly calculated `Aabb` -> `Sphere` and do `Sphere`-`Frustum` intersection tests before `Aabb`-`Frustum` tests. This is faster for `many_cubes` but could be slower in other cases where the sphere test gives a false-positive that the `Aabb` test discards. Also, I tested precalculating the `Sphere`s and inserting them alongside the `Aabb` but this was slower.
- Do not test meshes against the far plane. Apparently games don't do this anymore with infinite projections, and it's one fewer plane to test against. I made it optional and still do the test for culling lights but that is up for discussion.
- These collectively reduce `check_visibility` execution time in `many_cubes -- sphere` from 2.76ms to 1.48ms and increase frame rate from ~42fps to ~44fps
# Objective
fix cluster tilesize and tilecount calculations.
Fixes https://github.com/bevyengine/bevy/issues/4127 & https://github.com/bevyengine/bevy/issues/3596
## Solution
- calculate tilesize as smallest integers such that dimensions.xy() tiles will cover the screen
- calculate final dimensions as smallest integers such that final dimensions * tilesize will cover the screen
there is more cleanup that could be done in these functions. a future PR will likely remove the tilesize completely, so this is just a minimal change set to fix the current bug at small screen sizes
Co-authored-by: Carter Anderson <mcanders1@gmail.com>
# Objective
provide some customisation for default cluster setup
avoid "cluster index lists is full" in all cases (using a strategy outlined by @superdump)
## Solution
Add ClusterConfig enum (which can be inserted into a view at any time) to allow specifying cluster setup with variants:
- None (do not do any light assignment - for views which do not require light info, e.g. minimaps etc)
- Single (one cluster)
- XYZ (explicit cluster counts in each dimension)
- FixedZ (most similar to current - specify Z-slices and total, then x and y counts are dynamically determined to give approximately square clusters based on current aspect ratio)
Defaults to FixedZ { total: 4096, z: 24 } which is similar to the current setup.
Per frame, estimate the number of indices that would be required for the current config and decrease the cluster counts / increase the cluster sizes in the x and y dimensions if the index list would be too small.
notes:
- I didn't put ClusterConfig in the camera bundles to avoid introducing a dependency from bevy_render to bevy_pbr. the ClusterConfig enum comes with a pbr-centric impl block so i didn't want to move that into bevy_render either.
- ~Might want to add None variant to cluster config for views that don't care about lights?~
- Not well tested for orthographic
- ~there's a cluster_muck branch on my repo which includes some diagnostics / a modified lighting example which may be useful for tyre-kicking~ (outdated, i will bring it up to date if required)
anecdotal timings:
FPS on the lighting demo is negligibly better (~5%), maybe due to a small optimisation constraining the light aabb to be in front of the camera
FPS on the lighting demo with 100 extra lights added is ~33% faster, and also renders correctly as the cluster index count is no longer exceeded
## Objective
Currently, all directional and point lights have their viewing frusta recalculated every frame, even if they have not moved or been disabled/enabled.
## Solution
The relevant systems now make use of change detection to only update those lights whose viewing frusta may have changed.
# Objective
Add Visibility for lights
## Solution
- add Visibility to PointLightBundle and DirectionLightBundle
- filter lights used by Visibility.is_visible
note: includes changes from #3916 due to overlap, will be cleaner after that is merged
# Objective
fix#3915
## Solution
the issues are caused by
- lights are assigned to clusters before being filtered down to MAX_POINT_LIGHTS, leading to cluster counts potentially being too high
- after fixing the above, packing the count into 8 bits still causes overflow with exactly 256 lights affecting a cluster
to fix:
```assign_lights_to_clusters```
- limit extracted lights to MAX_POINT_LIGHTS, selecting based on shadow-caster & intensity (if required)
- warn if MAX_POINT_LIGHT count is exceeded
```prepare_lights```
- limit the lights assigned to a cluster to CLUSTER_COUNT_MASK (which is 1 less than MAX_POINT_LIGHTS) to avoid overflowing into the offset bits
notes:
- a better solution to the overflow may be to use more than 8 bits for cluster_count (the comment states only 14 of the remaining 24 bits are used for the offset). this would touch more of the code base but i'm happy to try if it has some benefit.
- intensity is only one way to select lights. it may be worth allowing user configuration of the light filtering, but i can't see a clean way to do that
# Objective
- Optimize assign_lights_to_clusters
## Solution
- Avoid inserting entities into hash sets in inner loops when it is known they will be inserted in at least one iteration of the loop.
- Use a Vec instead of a hash set where the set is not needed
- Avoid explicit calculation of the cluster_index from x,y,z coordinates, instead using row and column offsets and just adding z in the inner loop
- These changes cut the time spent in the system roughly in half
# Objective
Will fix#3377 and #3254
## Solution
Use an enum to represent either a `WindowId` or `Handle<Image>` in place of `Camera::window`.
Co-authored-by: Carter Anderson <mcanders1@gmail.com>
What is says on the tin.
This has got more to do with making `clippy` slightly more *quiet* than it does with changing anything that might greatly impact readability or performance.
that said, deriving `Default` for a couple of structs is a nice easy win
# Objective
The query for `VisiblePointLights` in `check_light_mesh_visibility` has a `Without<DirectionalLight>` filter. However, because `VisiblePointLights` is no longer an alias for `VisibleEntities`, the query won't conflict with the query for `DirectionalLight`s and thus the filter is unnecessary.
## Solution
Remove the filter and the outdated comment explaining its purpose.
# Objective
- Using plain exponential depth slicing for perspective projection cameras results in unnecessarily many slices very close together close to the camera. If the camera is then moved close to a collection of point lights, they will likely exhaust the available uniform buffer space for the lists of which lights affect which clusters.
## Solution
- A simple solution to this is to use a different near plane value for the depth slicing and set it to where the first slice's far plane should be. The default value is 5 and works well. This results in the configured number of depth slices, maintains the exponential slicing beyond the initial slice, and no slices are too small such that they cause problems that are sensitive to the view position.
# Objective
- Add support for loading lights from glTF 2.0 files
## Solution
- This adds support for the KHR_punctual_lights extension which supports point, directional, and spot lights, though we don't yet support spot lights.
- Inserting light bundles when creating scenes required registering some more light bundle component types.
# Objective
- 3d examples fail to run in webgl2 because of unsupported texture formats or texture too large
## Solution
- switch to supported formats if a feature is enabled. I choose a feature instead of a build target to not conflict with a potential webgpu support
Very inspired by 6813b2edc5, and need #3290 to work.
I named the feature `webgl2`, but it's only needed if one want to use PBR in webgl2. Examples using only 2D already work.
Co-authored-by: François <8672791+mockersf@users.noreply.github.com>
# Objective
Fixes: #3368
Issue was caused by screen size being: `(0, 0)`.
## Solution
Don't update clusters if the screen size is zero. A better solution might be to not render when minimized, but this works in the meantime.
Co-authored-by: John <startoaster23@gmail.com>
# Objective
This PR fixes a crash when winit is enabled when there is a camera in the world. Part of #3155
## Solution
In this PR, I removed two unwraps and added an example for regression testing.
# Objective
PBR lighting was broken in the new renderer when using orthographic projections due to the way the depth slicing works for the clusters. Fix it.
## Solution
- The default orthographic projection near plane is 0.0. The perspective projection depth slicing does a division by the near plane which gives a floating point NaN and the clustering all breaks down.
- Orthographic projections have a linear depth mapping, so it made intuitive sense to me to do depth slicing with a linear mapping too. The alternative I saw was to try to handle the near plane being at 0.0 and using the exponential depth slicing, but that felt like a hack that didn't make sense.
- As such, I have added code that detects whether the projection is orthographic based on `projection[3][3] == 1.0` and then implemented the orthographic mapping case throughout (when computing cluster AABBs, and when mapping a view space position (or light) to a cluster id in both the rust and shader code).
## Screenshots
Before:
After:
<img width="1392" alt="Screenshot 2021-12-13 at 16 36 53" src="https://user-images.githubusercontent.com/302146/145847314-6f3a2035-5d87-4896-8032-0c3e35e15b7d.png">
Old renderer (slightly lighter due to slight difference in configured intensity):
<img width="1392" alt="Screenshot 2021-12-13 at 16 42 23" src="https://user-images.githubusercontent.com/302146/145847391-6a5e6fe0-22da-4fc1-a6c7-440543689a63.png">
This makes the [New Bevy Renderer](#2535) the default (and only) renderer. The new renderer isn't _quite_ ready for the final release yet, but I want as many people as possible to start testing it so we can identify bugs and address feedback prior to release.
The examples are all ported over and operational with a few exceptions:
* I removed a good portion of the examples in the `shader` folder. We still have some work to do in order to make these examples possible / ergonomic / worthwhile: #3120 and "high level shader material plugins" are the big ones. This is a temporary measure.
* Temporarily removed the multiple_windows example: doing this properly in the new renderer will require the upcoming "render targets" changes. Same goes for the render_to_texture example.
* Removed z_sort_debug: entity visibility sort info is no longer available in app logic. we could do this on the "render app" side, but i dont consider it a priority.
This implements the most minimal variant of #1843 - a derive for marker trait. This is a prerequisite to more complicated features like statically defined storage type or opt-out component reflection.
In order to make component struct's purpose explicit and avoid misuse, it must be annotated with `#[derive(Component)]` (manual impl is discouraged for compatibility). Right now this is just a marker trait, but in the future it might be expanded. Making this change early allows us to make further changes later without breaking backward compatibility for derive macro users.
This already prevents a lot of issues, like using bundles in `insert` calls. Primitive types are no longer valid components as well. This can be easily worked around by adding newtype wrappers and deriving `Component` for them.
One funny example of prevented bad code (from our own tests) is when an newtype struct or enum variant is used. Previously, it was possible to write `insert(Newtype)` instead of `insert(Newtype(value))`. That code compiled, because function pointers (in this case newtype struct constructor) implement `Send + Sync + 'static`, so we allowed them to be used as components. This is no longer the case and such invalid code will trigger a compile error.
Co-authored-by: = <=>
Co-authored-by: TheRawMeatball <therawmeatball@gmail.com>
Co-authored-by: Carter Anderson <mcanders1@gmail.com>
This gets rid of multiple unsafe blocks that we had to maintain ourselves, and instead depends on library that's commonly used and supported by the ecosystem. We also get support for glam types for free.
There is still some things to clear up with the `Bytes` trait, but that is a bit more substantial change and can be done separately. Also there are already separate efforts to use `crevice` crate, so I've just added that as a TODO.
Introduced in #1778, not fixed by #1931
The size of `Lights` buffer currently is :
```rust
16 // (color, `[f32; 4]`)
+ 16 // (number of lights, `f32` encoded as a `[f32; 4]`)
+ 10 // (maximum number of lights)
* ( 16 // (light position, `[f32; 4]`
+ 16 // (color, `[16; 4]`)
+ 4 // (inverse_range_squared, `f32`)
)
-> 392
```
This makes the pbr shader crash when running with Xcode debugger or with the WebGL2 backend. They both expect a buffer sized 512. This can also be seen on desktop by adding a second light to a scene with a color, it's position and color will be wrong.
adding a second light to example `load_gltf`:
```rust
commands
.spawn_bundle(PointLightBundle {
transform: Transform::from_xyz(-3.0, 5.0, -3.0),
point_light: PointLight {
color: Color::BLUE,
..Default::default()
},
..Default::default()
})
.insert(Rotates);
```
before fix:
<img width="1392" alt="Screenshot 2021-04-16 at 19 14 59" src="https://user-images.githubusercontent.com/8672791/115060744-866fb080-9ee8-11eb-8915-f87cc872ad48.png">
after fix:
<img width="1392" alt="Screenshot 2021-04-16 at 19 16 44" src="https://user-images.githubusercontent.com/8672791/115060759-8cfe2800-9ee8-11eb-92c2-d79f39c7b36b.png">
This PR changes `inverse_range_squared` to be a `[f32; 4]` instead of a `f32` to have the expected alignement
After an inquiry on Reddit about support for Directional Lights and the unused properties on Light, I wanted to clean it up, to hopefully make it ever so slightly more clear for anyone wanting to add additional light types.
Co-authored-by: Carter Anderson <mcanders1@gmail.com>
This is a rebase of StarArawns PBR work from #261 with IngmarBitters work from #1160 cherry-picked on top.
I had to make a few minor changes to make some intermediate commits compile and the end result is not yet 100% what I expected, so there's a bit more work to do.
Co-authored-by: John Mitchell <toasterthegamer@gmail.com>
Co-authored-by: Ingmar Bitter <ingmar.bitter@gmail.com>
Idea being this would be easier to grasp for end-users. Problem with the logical defaults is this breaks current setups, because light will become 20 times less bright. But most folks won't have customized this resource or will not have used `..Default::default()` due to lack of other fields.
# Bevy ECS V2
This is a rewrite of Bevy ECS (basically everything but the new executor/schedule, which are already awesome). The overall goal was to improve the performance and versatility of Bevy ECS. Here is a quick bulleted list of changes before we dive into the details:
* Complete World rewrite
* Multiple component storage types:
* Tables: fast cache friendly iteration, slower add/removes (previously called Archetypes)
* Sparse Sets: fast add/remove, slower iteration
* Stateful Queries (caches query results for faster iteration. fragmented iteration is _fast_ now)
* Stateful System Params (caches expensive operations. inspired by @DJMcNab's work in #1364)
* Configurable System Params (users can set configuration when they construct their systems. once again inspired by @DJMcNab's work)
* Archetypes are now "just metadata", component storage is separate
* Archetype Graph (for faster archetype changes)
* Component Metadata
* Configure component storage type
* Retrieve information about component size/type/name/layout/send-ness/etc
* Components are uniquely identified by a densely packed ComponentId
* TypeIds are now totally optional (which should make implementing scripting easier)
* Super fast "for_each" query iterators
* Merged Resources into World. Resources are now just a special type of component
* EntityRef/EntityMut builder apis (more efficient and more ergonomic)
* Fast bitset-backed `Access<T>` replaces old hashmap-based approach everywhere
* Query conflicts are determined by component access instead of archetype component access (to avoid random failures at runtime)
* With/Without are still taken into account for conflicts, so this should still be comfy to use
* Much simpler `IntoSystem` impl
* Significantly reduced the amount of hashing throughout the ecs in favor of Sparse Sets (indexed by densely packed ArchetypeId, ComponentId, BundleId, and TableId)
* Safety Improvements
* Entity reservation uses a normal world reference instead of unsafe transmute
* QuerySets no longer transmute lifetimes
* Made traits "unsafe" where relevant
* More thorough safety docs
* WorldCell
* Exposes safe mutable access to multiple resources at a time in a World
* Replaced "catch all" `System::update_archetypes(world: &World)` with `System::new_archetype(archetype: &Archetype)`
* Simpler Bundle implementation
* Replaced slow "remove_bundle_one_by_one" used as fallback for Commands::remove_bundle with fast "remove_bundle_intersection"
* Removed `Mut<T>` query impl. it is better to only support one way: `&mut T`
* Removed with() from `Flags<T>` in favor of `Option<Flags<T>>`, which allows querying for flags to be "filtered" by default
* Components now have is_send property (currently only resources support non-send)
* More granular module organization
* New `RemovedComponents<T>` SystemParam that replaces `query.removed::<T>()`
* `world.resource_scope()` for mutable access to resources and world at the same time
* WorldQuery and QueryFilter traits unified. FilterFetch trait added to enable "short circuit" filtering. Auto impled for cases that don't need it
* Significantly slimmed down SystemState in favor of individual SystemParam state
* System Commands changed from `commands: &mut Commands` back to `mut commands: Commands` (to allow Commands to have a World reference)
Fixes#1320
## `World` Rewrite
This is a from-scratch rewrite of `World` that fills the niche that `hecs` used to. Yes, this means Bevy ECS is no longer a "fork" of hecs. We're going out our own!
(the only shared code between the projects is the entity id allocator, which is already basically ideal)
A huge shout out to @SanderMertens (author of [flecs](https://github.com/SanderMertens/flecs)) for sharing some great ideas with me (specifically hybrid ecs storage and archetype graphs). He also helped advise on a number of implementation details.
## Component Storage (The Problem)
Two ECS storage paradigms have gained a lot of traction over the years:
* **Archetypal ECS**:
* Stores components in "tables" with static schemas. Each "column" stores components of a given type. Each "row" is an entity.
* Each "archetype" has its own table. Adding/removing an entity's component changes the archetype.
* Enables super-fast Query iteration due to its cache-friendly data layout
* Comes at the cost of more expensive add/remove operations for an Entity's components, because all components need to be copied to the new archetype's "table"
* **Sparse Set ECS**:
* Stores components of the same type in densely packed arrays, which are sparsely indexed by densely packed unsigned integers (Entity ids)
* Query iteration is slower than Archetypal ECS because each entity's component could be at any position in the sparse set. This "random access" pattern isn't cache friendly. Additionally, there is an extra layer of indirection because you must first map the entity id to an index in the component array.
* Adding/removing components is a cheap, constant time operation
Bevy ECS V1, hecs, legion, flec, and Unity DOTS are all "archetypal ecs-es". I personally think "archetypal" storage is a good default for game engines. An entity's archetype doesn't need to change frequently in general, and it creates "fast by default" query iteration (which is a much more common operation). It is also "self optimizing". Users don't need to think about optimizing component layouts for iteration performance. It "just works" without any extra boilerplate.
Shipyard and EnTT are "sparse set ecs-es". They employ "packing" as a way to work around the "suboptimal by default" iteration performance for specific sets of components. This helps, but I didn't think this was a good choice for a general purpose engine like Bevy because:
1. "packs" conflict with each other. If bevy decides to internally pack the Transform and GlobalTransform components, users are then blocked if they want to pack some custom component with Transform.
2. users need to take manual action to optimize
Developers selecting an ECS framework are stuck with a hard choice. Select an "archetypal" framework with "fast iteration everywhere" but without the ability to cheaply add/remove components, or select a "sparse set" framework to cheaply add/remove components but with slower iteration performance.
## Hybrid Component Storage (The Solution)
In Bevy ECS V2, we get to have our cake and eat it too. It now has _both_ of the component storage types above (and more can be added later if needed):
* **Tables** (aka "archetypal" storage)
* The default storage. If you don't configure anything, this is what you get
* Fast iteration by default
* Slower add/remove operations
* **Sparse Sets**
* Opt-in
* Slower iteration
* Faster add/remove operations
These storage types complement each other perfectly. By default Query iteration is fast. If developers know that they want to add/remove a component at high frequencies, they can set the storage to "sparse set":
```rust
world.register_component(
ComponentDescriptor:🆕:<MyComponent>(StorageType::SparseSet)
).unwrap();
```
## Archetypes
Archetypes are now "just metadata" ... they no longer store components directly. They do store:
* The `ComponentId`s of each of the Archetype's components (and that component's storage type)
* Archetypes are uniquely defined by their component layouts
* For example: entities with "table" components `[A, B, C]` _and_ "sparse set" components `[D, E]` will always be in the same archetype.
* The `TableId` associated with the archetype
* For now each archetype has exactly one table (which can have no components),
* There is a 1->Many relationship from Tables->Archetypes. A given table could have any number of archetype components stored in it:
* Ex: an entity with "table storage" components `[A, B, C]` and "sparse set" components `[D, E]` will share the same `[A, B, C]` table as an entity with `[A, B, C]` table component and `[F]` sparse set components.
* This 1->Many relationship is how we preserve fast "cache friendly" iteration performance when possible (more on this later)
* A list of entities that are in the archetype and the row id of the table they are in
* ArchetypeComponentIds
* unique densely packed identifiers for (ArchetypeId, ComponentId) pairs
* used by the schedule executor for cheap system access control
* "Archetype Graph Edges" (see the next section)
## The "Archetype Graph"
Archetype changes in Bevy (and a number of other archetypal ecs-es) have historically been expensive to compute. First, you need to allocate a new vector of the entity's current component ids, add or remove components based on the operation performed, sort it (to ensure it is order-independent), then hash it to find the archetype (if it exists). And thats all before we get to the _already_ expensive full copy of all components to the new table storage.
The solution is to build a "graph" of archetypes to cache these results. @SanderMertens first exposed me to the idea (and he got it from @gjroelofs, who came up with it). They propose adding directed edges between archetypes for add/remove component operations. If `ComponentId`s are densely packed, you can use sparse sets to cheaply jump between archetypes.
Bevy takes this one step further by using add/remove `Bundle` edges instead of `Component` edges. Bevy encourages the use of `Bundles` to group add/remove operations. This is largely for "clearer game logic" reasons, but it also helps cut down on the number of archetype changes required. `Bundles` now also have densely-packed `BundleId`s. This allows us to use a _single_ edge for each bundle operation (rather than needing to traverse N edges ... one for each component). Single component operations are also bundles, so this is strictly an improvement over a "component only" graph.
As a result, an operation that used to be _heavy_ (both for allocations and compute) is now two dirt-cheap array lookups and zero allocations.
## Stateful Queries
World queries are now stateful. This allows us to:
1. Cache archetype (and table) matches
* This resolves another issue with (naive) archetypal ECS: query performance getting worse as the number of archetypes goes up (and fragmentation occurs).
2. Cache Fetch and Filter state
* The expensive parts of fetch/filter operations (such as hashing the TypeId to find the ComponentId) now only happen once when the Query is first constructed
3. Incrementally build up state
* When new archetypes are added, we only process the new archetypes (no need to rebuild state for old archetypes)
As a result, the direct `World` query api now looks like this:
```rust
let mut query = world.query::<(&A, &mut B)>();
for (a, mut b) in query.iter_mut(&mut world) {
}
```
Requiring `World` to generate stateful queries (rather than letting the `QueryState` type be constructed separately) allows us to ensure that _all_ queries are properly initialized (and the relevant world state, such as ComponentIds). This enables QueryState to remove branches from its operations that check for initialization status (and also enables query.iter() to take an immutable world reference because it doesn't need to initialize anything in world).
However in systems, this is a non-breaking change. State management is done internally by the relevant SystemParam.
## Stateful SystemParams
Like Queries, `SystemParams` now also cache state. For example, `Query` system params store the "stateful query" state mentioned above. Commands store their internal `CommandQueue`. This means you can now safely use as many separate `Commands` parameters in your system as you want. `Local<T>` system params store their `T` value in their state (instead of in Resources).
SystemParam state also enabled a significant slim-down of SystemState. It is much nicer to look at now.
Per-SystemParam state naturally insulates us from an "aliased mut" class of errors we have hit in the past (ex: using multiple `Commands` system params).
(credit goes to @DJMcNab for the initial idea and draft pr here #1364)
## Configurable SystemParams
@DJMcNab also had the great idea to make SystemParams configurable. This allows users to provide some initial configuration / values for system parameters (when possible). Most SystemParams have no config (the config type is `()`), but the `Local<T>` param now supports user-provided parameters:
```rust
fn foo(value: Local<usize>) {
}
app.add_system(foo.system().config(|c| c.0 = Some(10)));
```
## Uber Fast "for_each" Query Iterators
Developers now have the choice to use a fast "for_each" iterator, which yields ~1.5-3x iteration speed improvements for "fragmented iteration", and minor ~1.2x iteration speed improvements for unfragmented iteration.
```rust
fn system(query: Query<(&A, &mut B)>) {
// you now have the option to do this for a speed boost
query.for_each_mut(|(a, mut b)| {
});
// however normal iterators are still available
for (a, mut b) in query.iter_mut() {
}
}
```
I think in most cases we should continue to encourage "normal" iterators as they are more flexible and more "rust idiomatic". But when that extra "oomf" is needed, it makes sense to use `for_each`.
We should also consider using `for_each` for internal bevy systems to give our users a nice speed boost (but that should be a separate pr).
## Component Metadata
`World` now has a `Components` collection, which is accessible via `world.components()`. This stores mappings from `ComponentId` to `ComponentInfo`, as well as `TypeId` to `ComponentId` mappings (where relevant). `ComponentInfo` stores information about the component, such as ComponentId, TypeId, memory layout, send-ness (currently limited to resources), and storage type.
## Significantly Cheaper `Access<T>`
We used to use `TypeAccess<TypeId>` to manage read/write component/archetype-component access. This was expensive because TypeIds must be hashed and compared individually. The parallel executor got around this by "condensing" type ids into bitset-backed access types. This worked, but it had to be re-generated from the `TypeAccess<TypeId>`sources every time archetypes changed.
This pr removes TypeAccess in favor of faster bitset access everywhere. We can do this thanks to the move to densely packed `ComponentId`s and `ArchetypeComponentId`s.
## Merged Resources into World
Resources had a lot of redundant functionality with Components. They stored typed data, they had access control, they had unique ids, they were queryable via SystemParams, etc. In fact the _only_ major difference between them was that they were unique (and didn't correlate to an entity).
Separate resources also had the downside of requiring a separate set of access controls, which meant the parallel executor needed to compare more bitsets per system and manage more state.
I initially got the "separate resources" idea from `legion`. I think that design was motivated by the fact that it made the direct world query/resource lifetime interactions more manageable. It certainly made our lives easier when using Resources alongside hecs/bevy_ecs. However we already have a construct for safely and ergonomically managing in-world lifetimes: systems (which use `Access<T>` internally).
This pr merges Resources into World:
```rust
world.insert_resource(1);
world.insert_resource(2.0);
let a = world.get_resource::<i32>().unwrap();
let mut b = world.get_resource_mut::<f64>().unwrap();
*b = 3.0;
```
Resources are now just a special kind of component. They have their own ComponentIds (and their own resource TypeId->ComponentId scope, so they don't conflict wit components of the same type). They are stored in a special "resource archetype", which stores components inside the archetype using a new `unique_components` sparse set (note that this sparse set could later be used to implement Tags). This allows us to keep the code size small by reusing existing datastructures (namely Column, Archetype, ComponentFlags, and ComponentInfo). This allows us the executor to use a single `Access<ArchetypeComponentId>` per system. It should also make scripting language integration easier.
_But_ this merge did create problems for people directly interacting with `World`. What if you need mutable access to multiple resources at the same time? `world.get_resource_mut()` borrows World mutably!
## WorldCell
WorldCell applies the `Access<ArchetypeComponentId>` concept to direct world access:
```rust
let world_cell = world.cell();
let a = world_cell.get_resource_mut::<i32>().unwrap();
let b = world_cell.get_resource_mut::<f64>().unwrap();
```
This adds cheap runtime checks (a sparse set lookup of `ArchetypeComponentId` and a counter) to ensure that world accesses do not conflict with each other. Each operation returns a `WorldBorrow<'w, T>` or `WorldBorrowMut<'w, T>` wrapper type, which will release the relevant ArchetypeComponentId resources when dropped.
World caches the access sparse set (and only one cell can exist at a time), so `world.cell()` is a cheap operation.
WorldCell does _not_ use atomic operations. It is non-send, does a mutable borrow of world to prevent other accesses, and uses a simple `Rc<RefCell<ArchetypeComponentAccess>>` wrapper in each WorldBorrow pointer.
The api is currently limited to resource access, but it can and should be extended to queries / entity component access.
## Resource Scopes
WorldCell does not yet support component queries, and even when it does there are sometimes legitimate reasons to want a mutable world ref _and_ a mutable resource ref (ex: bevy_render and bevy_scene both need this). In these cases we could always drop down to the unsafe `world.get_resource_unchecked_mut()`, but that is not ideal!
Instead developers can use a "resource scope"
```rust
world.resource_scope(|world: &mut World, a: &mut A| {
})
```
This temporarily removes the `A` resource from `World`, provides mutable pointers to both, and re-adds A to World when finished. Thanks to the move to ComponentIds/sparse sets, this is a cheap operation.
If multiple resources are required, scopes can be nested. We could also consider adding a "resource tuple" to the api if this pattern becomes common and the boilerplate gets nasty.
## Query Conflicts Use ComponentId Instead of ArchetypeComponentId
For safety reasons, systems cannot contain queries that conflict with each other without wrapping them in a QuerySet. On bevy `main`, we use ArchetypeComponentIds to determine conflicts. This is nice because it can take into account filters:
```rust
// these queries will never conflict due to their filters
fn filter_system(a: Query<&mut A, With<B>>, b: Query<&mut B, Without<B>>) {
}
```
But it also has a significant downside:
```rust
// these queries will not conflict _until_ an entity with A, B, and C is spawned
fn maybe_conflicts_system(a: Query<(&mut A, &C)>, b: Query<(&mut A, &B)>) {
}
```
The system above will panic at runtime if an entity with A, B, and C is spawned. This makes it hard to trust that your game logic will run without crashing.
In this pr, I switched to using `ComponentId` instead. This _is_ more constraining. `maybe_conflicts_system` will now always fail, but it will do it consistently at startup. Naively, it would also _disallow_ `filter_system`, which would be a significant downgrade in usability. Bevy has a number of internal systems that rely on disjoint queries and I expect it to be a common pattern in userspace.
To resolve this, I added a new `FilteredAccess<T>` type, which wraps `Access<T>` and adds with/without filters. If two `FilteredAccess` have with/without values that prove they are disjoint, they will no longer conflict.
## EntityRef / EntityMut
World entity operations on `main` require that the user passes in an `entity` id to each operation:
```rust
let entity = world.spawn((A, )); // create a new entity with A
world.get::<A>(entity);
world.insert(entity, (B, C));
world.insert_one(entity, D);
```
This means that each operation needs to look up the entity location / verify its validity. The initial spawn operation also requires a Bundle as input. This can be awkward when no components are required (or one component is required).
These operations have been replaced by `EntityRef` and `EntityMut`, which are "builder-style" wrappers around world that provide read and read/write operations on a single, pre-validated entity:
```rust
// spawn now takes no inputs and returns an EntityMut
let entity = world.spawn()
.insert(A) // insert a single component into the entity
.insert_bundle((B, C)) // insert a bundle of components into the entity
.id() // id returns the Entity id
// Returns EntityMut (or panics if the entity does not exist)
world.entity_mut(entity)
.insert(D)
.insert_bundle(SomeBundle::default());
{
// returns EntityRef (or panics if the entity does not exist)
let d = world.entity(entity)
.get::<D>() // gets the D component
.unwrap();
// world.get still exists for ergonomics
let d = world.get::<D>(entity).unwrap();
}
// These variants return Options if you want to check existence instead of panicing
world.get_entity_mut(entity)
.unwrap()
.insert(E);
if let Some(entity_ref) = world.get_entity(entity) {
let d = entity_ref.get::<D>().unwrap();
}
```
This _does not_ affect the current Commands api or terminology. I think that should be a separate conversation as that is a much larger breaking change.
## Safety Improvements
* Entity reservation in Commands uses a normal world borrow instead of an unsafe transmute
* QuerySets no longer transmutes lifetimes
* Made traits "unsafe" when implementing a trait incorrectly could cause unsafety
* More thorough safety docs
## RemovedComponents SystemParam
The old approach to querying removed components: `query.removed:<T>()` was confusing because it had no connection to the query itself. I replaced it with the following, which is both clearer and allows us to cache the ComponentId mapping in the SystemParamState:
```rust
fn system(removed: RemovedComponents<T>) {
for entity in removed.iter() {
}
}
```
## Simpler Bundle implementation
Bundles are no longer responsible for sorting (or deduping) TypeInfo. They are just a simple ordered list of component types / data. This makes the implementation smaller and opens the door to an easy "nested bundle" implementation in the future (which i might even add in this pr). Duplicate detection is now done once per bundle type by World the first time a bundle is used.
## Unified WorldQuery and QueryFilter types
(don't worry they are still separate type _parameters_ in Queries .. this is a non-breaking change)
WorldQuery and QueryFilter were already basically identical apis. With the addition of `FetchState` and more storage-specific fetch methods, the overlap was even clearer (and the redundancy more painful).
QueryFilters are now just `F: WorldQuery where F::Fetch: FilterFetch`. FilterFetch requires `Fetch<Item = bool>` and adds new "short circuit" variants of fetch methods. This enables a filter tuple like `(With<A>, Without<B>, Changed<C>)` to stop evaluating the filter after the first mismatch is encountered. FilterFetch is automatically implemented for `Fetch` implementations that return bool.
This forces fetch implementations that return things like `(bool, bool, bool)` (such as the filter above) to manually implement FilterFetch and decide whether or not to short-circuit.
## More Granular Modules
World no longer globs all of the internal modules together. It now exports `core`, `system`, and `schedule` separately. I'm also considering exporting `core` submodules directly as that is still pretty "glob-ey" and unorganized (feedback welcome here).
## Remaining Draft Work (to be done in this pr)
* ~~panic on conflicting WorldQuery fetches (&A, &mut A)~~
* ~~bevy `main` and hecs both currently allow this, but we should protect against it if possible~~
* ~~batch_iter / par_iter (currently stubbed out)~~
* ~~ChangedRes~~
* ~~I skipped this while we sort out #1313. This pr should be adapted to account for whatever we land on there~~.
* ~~The `Archetypes` and `Tables` collections use hashes of sorted lists of component ids to uniquely identify each archetype/table. This hash is then used as the key in a HashMap to look up the relevant ArchetypeId or TableId. (which doesn't handle hash collisions properly)~~
* ~~It is currently unsafe to generate a Query from "World A", then use it on "World B" (despite the api claiming it is safe). We should probably close this gap. This could be done by adding a randomly generated WorldId to each world, then storing that id in each Query. They could then be compared to each other on each `query.do_thing(&world)` operation. This _does_ add an extra branch to each query operation, so I'm open to other suggestions if people have them.~~
* ~~Nested Bundles (if i find time)~~
## Potential Future Work
* Expand WorldCell to support queries.
* Consider not allocating in the empty archetype on `world.spawn()`
* ex: return something like EntityMutUninit, which turns into EntityMut after an `insert` or `insert_bundle` op
* this actually regressed performance last time i tried it, but in theory it should be faster
* Optimize SparseSet::insert (see `PERF` comment on insert)
* Replace SparseArray `Option<T>` with T::MAX to cut down on branching
* would enable cheaper get_unchecked() operations
* upstream fixedbitset optimizations
* fixedbitset could be allocation free for small block counts (store blocks in a SmallVec)
* fixedbitset could have a const constructor
* Consider implementing Tags (archetype-specific by-value data that affects archetype identity)
* ex: ArchetypeA could have `[A, B, C]` table components and `[D(1)]` "tag" component. ArchetypeB could have `[A, B, C]` table components and a `[D(2)]` tag component. The archetypes are different, despite both having D tags because the value inside D is different.
* this could potentially build on top of the `archetype.unique_components` added in this pr for resource storage.
* Consider reverting `all_tuples` proc macro in favor of the old `macro_rules` implementation
* all_tuples is more flexible and produces cleaner documentation (the macro_rules version produces weird type parameter orders due to parser constraints)
* but unfortunately all_tuples also appears to make Rust Analyzer sad/slow when working inside of `bevy_ecs` (does not affect user code)
* Consider "resource queries" and/or "mixed resource and entity component queries" as an alternative to WorldCell
* this is basically just "systems" so maybe it's not worth it
* Add more world ops
* `world.clear()`
* `world.reserve<T: Bundle>(count: usize)`
* Try using the old archetype allocation strategy (allocate new memory on resize and copy everything over). I expect this to improve batch insertion performance at the cost of unbatched performance. But thats just a guess. I'm not an allocation perf pro :)
* Adapt Commands apis for consistency with new World apis
## Benchmarks
key:
* `bevy_old`: bevy `main` branch
* `bevy`: this branch
* `_foreach`: uses an optimized for_each iterator
* ` _sparse`: uses sparse set storage (if unspecified assume table storage)
* `_system`: runs inside a system (if unspecified assume test happens via direct world ops)
### Simple Insert (from ecs_bench_suite)
### Simpler Iter (from ecs_bench_suite)
### Fragment Iter (from ecs_bench_suite)
### Sparse Fragmented Iter
Iterate a query that matches 5 entities from a single matching archetype, but there are 100 unmatching archetypes
### Schedule (from ecs_bench_suite)
### Add Remove Component (from ecs_bench_suite)
### Add Remove Component Big
Same as the test above, but each entity has 5 "large" matrix components and 1 "large" matrix component is added and removed
### Get Component
Looks up a single component value a large number of times
Transform and GlobalTransform are now Similarities.
This resolves precision errors and simplifies the api
Co-authored-by: Carter Anderson <mcanders1@gmail.com>
