I didn't notice minus where vertices are generated, so could not
understand the order there.
Adding a comment to help the next person who is going to understand Bevy
by reading its code.
# Objective
- Reduce nesting in `process_handle_drop_internal()`.
- Related to #10896.
## Solution
- Use early returns when possible.
- Reduced from 9 levels of indents to 4.
# Objective
add `RenderLayers` awareness to lights. lights default to
`RenderLayers::layer(0)`, and must intersect the camera entity's
`RenderLayers` in order to affect the camera's output.
note that lights already use renderlayers to filter meshes for shadow
casting. this adds filtering lights per view based on intersection of
camera layers and light layers.
fixes#3462
## Solution
PointLights and SpotLights are assigned to individual views in
`assign_lights_to_clusters`, so we simply cull the lights which don't
match the view layers in that function.
DirectionalLights are global, so we
- add the light layers to the `DirectionalLight` struct
- add the view layers to the `ViewUniform` struct
- check for intersection before processing the light in
`apply_pbr_lighting`
potential issue: when mesh/light layers are smaller than the view layers
weird results can occur. e.g:
camera = layers 1+2
light = layers 1
mesh = layers 2
the mesh does not cast shadows wrt the light as (1 & 2) == 0.
the light affects the view as (1+2 & 1) != 0.
the view renders the mesh as (1+2 & 2) != 0.
so the mesh is rendered and lit, but does not cast a shadow.
this could be fixed (so that the light would not affect the mesh in that
view) by adding the light layers to the point and spot light structs,
but i think the setup is pretty unusual, and space is at a premium in
those structs (adding 4 bytes more would reduce the webgl point+spot
light max count to 240 from 256).
I think typical usage is for cameras to have a single layer, and
meshes/lights to maybe have multiple layers to render to e.g. minimaps
as well as primary views.
if there is a good use case for the above setup and we should support
it, please let me know.
---
## Migration Guide
Lights no longer affect all `RenderLayers` by default, now like cameras
and meshes they default to `RenderLayers::layer(0)`. To recover the
previous behaviour and have all lights affect all views, add a
`RenderLayers::all()` component to the light entity.
# Objective
Printing `DynamicStruct` with a debug format does not show the contained
type anymore. For instance, in `examples/reflection/reflection.rs`,
adding `dbg!(&reflect_value);` to line 96 will print:
```rust
[examples/reflection/reflection.rs:96] &reflect_value = DynamicStruct(bevy_reflect::DynamicStruct {
a: 4,
nested: DynamicStruct(bevy_reflect::DynamicStruct {
b: 8,
}),
})
```
## Solution
Show the represented type instead (`reflection::Foo` and
`reflection::Bar` in this case):
```rust
[examples/reflection/reflection.rs:96] &reflect_value = DynamicStruct(reflection::Foo {
a: 4,
nested: DynamicStruct(reflection::Bar {
b: 8,
}),
})
```
---------
Co-authored-by: Gino Valente <49806985+MrGVSV@users.noreply.github.com>
# Objective
The `Despawn` command breaks the hierarchy whenever you use it if the
despawned entity has a parent or any children. This is a serious footgun
because the `Despawn` command has the shortest name, the behavior is
unexpected and not likely to be what you want, and the crash that it
causes can be very difficult to track down.
## Solution
Until this can be fixed by relations, add a note mentioning the footgun
in the documentation.
## Solution
`Commands.remove` and `.retain` (because I copied `remove`s doc)
referenced `EntityWorldMut.remove` and `retain` for more detail but the
`Commands` docs are much more detailed (which makes sense because it is
the most common api), so I have instead inverted this so that
`EntityWorldMut` docs link to `Commands`.
I also made `EntityWorldMut.despawn` reference `World.despawn` for more
details, like `Commands.despawn` does.
# Objective
Fixes#5891.
For mikktspace normal maps, normals must be renormalized in vertex
shaders to match the way mikktspace bakes vertex tangents and normal
maps so that the exact inverse process is applied when shading.
However, for invalid normals like `vec3<f32>(0.0, 0.0, 0.0)`, this
normalization causes NaN values, and because it's in the vertex shader,
it affects the entire triangle and causes it to be shaded as black:
*A cone with a tip that has a vertex normal of [0, 0, 0], causing the
mesh to be shaded as black.*
In some cases, normals of zero are actually *useful*. For example, a
smoothly shaded cone without creases requires the apex vertex normal to
be zero, because there is no singular normal that works correctly, so
the apex shouldn't contribute to the overall shading. Duplicate vertices
for the apex fix some shading issues, but it causes visible creases and
is more expensive. See #5891 and #10298 for more details.
For correctly shaded cones and other similar low-density shapes with
sharp tips, vertex normals of zero can not be normalized in the vertex
shader.
## Solution
Only normalize the vertex normals and tangents in the vertex shader if
the normal isn't [0, 0, 0]. This way, mikktspace normal maps should
still work for everything except the zero normals, and the zero normals
will only be normalized in the fragment shader.
This allows us to render cones correctly:
Notice how there is still a weird shadow banding effect in one area. I
noticed that it can be fixed by normalizing
[here](d2614f2d80/crates/bevy_pbr/src/render/pbr_functions.wgsl (L51)),
which produces a perfectly smooth cone without duplicate vertices:
I didn't add this change yet, because it seems a bit arbitrary. I can
add it here if that'd be useful or make another PR though.
# Objective
`update_accessibility_nodes` is one of the most nested functions in the
entire Bevy repository, with a maximum of 9 levels of indentations. This
PR refactors it down to 3 levels of indentations, while improving
readability on other fronts. The result is a function that is actually
understandable at a first glance.
- This is a proof of concept to demonstrate that it is possible to
gradually lower the nesting limit proposed by #10896.
PS: I read AccessKit's documentation, but I don't have experience with
it. Therefore, naming of variables and subroutines may be a bit off.
PS2: I don't know if the test suite covers the functionality of this
system, but since I've spent quite some time on it and the changes were
simple, I'm pretty confident the refactor is functionally equivalent to
the original.
## Solution
I documented each change with a commit, but as a summary I did the
following to reduce nesting:
- I moved from `if-let` blocks to `let-else` statements where
appropriate to reduce rightward shift
- I extracted the closure body to a new function `update_adapter`
- I factored out parts of `update_adapter` into new functions
`queue_node_for_update` and `add_children_nodes`
**Note for reviewers:** GitHub's diff viewer is not the greatest in
showing horizontal code shifts, therefore you may want to use a
different tool like VSCode to review some commits, especially the second
one (anyway, that commit is very straightforward, despite changing many
lines).
# Objective
A nodes outline should be clipped using its own clipping rect, not its
parents.
fixes#10921
## Solution
Clip outlines by the node's own clipping rect, not the parent's.
If you compare the `overflow` ui example in main with this PR, you'll
see that the outlines that appear when you hover above the images are
now clipped along with the images.
---
## Changelog
* Outlines are now clipped using the node's own clipping rect, not the
parent's.
# Objective
- Improve readability.
- Somewhat relates to #10896.
## Solution
- Use early returns to minimize nesting.
- Change `emitter_translation` to use `if let` instead of `map`.
# Objective
The `update_emitter_positions`, and `update_listener_positions` systems
are added for every call to `add_audio_source`.
Instead, add them once in the `AudioPlugin` directly.
Also merged the calls to `add_systems`.
Caught while working on my schedule visualizer c:
# Objective
Test more complex function signatures for exclusive systems, and test
that `StaticSystemParam` is indeed a `SystemParam`.
I mean, it currently works, but might as well add a test for it.
# Objective
- Resolves#10853
## Solution
- ~~Changed the name of `Input` struct to `PressableInput`.~~
- Changed the name of `Input` struct to `ButtonInput`.
## Migration Guide
- Breaking Change: Users need to rename `Input` to `ButtonInput` in
their projects.
# Objective
A better alternative version of #10843.
Currently, Bevy has a single `Ray` struct for 3D. To allow better
interoperability with Bevy's primitive shapes (#10572) and some third
party crates (that handle e.g. spatial queries), it would be very useful
to have separate versions for 2D and 3D respectively.
## Solution
Separate `Ray` into `Ray2d` and `Ray3d`. These new structs also take
advantage of the new primitives by using `Direction2d`/`Direction3d` for
the direction:
```rust
pub struct Ray2d {
pub origin: Vec2,
pub direction: Direction2d,
}
pub struct Ray3d {
pub origin: Vec3,
pub direction: Direction3d,
}
```
and by using `Plane2d`/`Plane3d` in `intersect_plane`:
```rust
impl Ray2d {
// ...
pub fn intersect_plane(&self, plane_origin: Vec2, plane: Plane2d) -> Option<f32> {
// ...
}
}
```
---
## Changelog
### Added
- `Ray2d` and `Ray3d`
- `Ray2d::new` and `Ray3d::new` constructors
- `Plane2d::new` and `Plane3d::new` constructors
### Removed
- Removed `Ray` in favor of `Ray3d`
### Changed
- `direction` is now a `Direction2d`/`Direction3d` instead of a vector,
which provides guaranteed normalization
- `intersect_plane` now takes a `Plane2d`/`Plane3d` instead of just a
vector for the plane normal
- `Direction2d` and `Direction3d` now derive `Serialize` and
`Deserialize` to preserve ray (de)serialization
## Migration Guide
`Ray` has been renamed to `Ray3d`.
### Ray creation
Before:
```rust
Ray {
origin: Vec3::ZERO,
direction: Vec3::new(0.5, 0.6, 0.2).normalize(),
}
```
After:
```rust
// Option 1:
Ray3d {
origin: Vec3::ZERO,
direction: Direction3d::new(Vec3::new(0.5, 0.6, 0.2)).unwrap(),
}
// Option 2:
Ray3d::new(Vec3::ZERO, Vec3::new(0.5, 0.6, 0.2))
```
### Plane intersections
Before:
```rust
let result = ray.intersect_plane(Vec2::X, Vec2::Y);
```
After:
```rust
let result = ray.intersect_plane(Vec2::X, Plane2d::new(Vec2::Y));
```
# Objective
Implement `TryFrom<Vec2>`/`TryFrom<Vec3>` for direction primitives as
considered in #10857.
## Solution
Implement `TryFrom` for the direction primitives.
These are all equivalent:
```rust
let dir2d = Direction2d::try_from(Vec2::new(0.5, 0.5)).unwrap();
let dir2d = Vec2::new(0.5, 0.5).try_into().unwrap(); // (assumes that the type is inferred)
let dir2d = Direction2d::new(Vec2::new(0.5, 0.5)).unwrap();
```
For error cases, an `Err(InvalidDirectionError)` is returned. It
contains the type of failure:
```rust
/// An error indicating that a direction is invalid.
#[derive(Debug, PartialEq)]
pub enum InvalidDirectionError {
/// The length of the direction vector is zero or very close to zero.
Zero,
/// The length of the direction vector is `std::f32::INFINITY`.
Infinite,
/// The length of the direction vector is `NaN`.
NaN,
}
```
# Objective
- Resolves#10784
## Solution
- As @ickshonpe mentioned in #10784, this is intended behavior but could
benefit from mentioning it in docs.
- I'm also thinking about adding a helper function to disable os scaling
such as `disable_os_scaling()`, but not sure if it's needed.
---
## Changelog
> Add a comment about scaling behavior that happens, and point user to
how he can avoid that behavior.
# Objective
Adds `EntityCommands.retain` and `EntityWorldMut.retain` to remove all
components except the given bundle from the entity.
Fixes#10865.
## Solution
I added a private unsafe function in `EntityWorldMut` called
`remove_bundle_info` which performs the shared behaviour of `remove` and
`retain`, namely taking a `BundleInfo` of components to remove, and
removing them from the given entity. Then `retain` simply gets all the
components on the entity and filters them by whether they are in the
bundle it was passed, before passing this `BundleInfo` into
`remove_bundle_info`.
`EntityCommands.retain` just creates a new type `Retain` which runs
`EntityWorldMut.retain` when run.
---
## Changelog
Added `EntityCommands.retain` and `EntityWorldMut.retain`, which remove
all components except the given bundle from the entity, they can also be
used to remove all components by passing `()` as the bundle.
# Objective
The name `TextAlignment` is really deceptive and almost every new user
gets confused about the differences between aligning text with
`TextAlignment`, aligning text with `Style` and aligning text with
anchor (when using `Text2d`).
## Solution
* Rename `TextAlignment` to `JustifyText`. The associated helper methods
are also renamed.
* Improve the doc comments for text explaining explicitly how the
`JustifyText` component affects the arrangement of text.
* Add some extra cases to the `text_debug` example that demonstate the
differences between alignment using `JustifyText` and alignment using
`Style`.
<img width="757" alt="text_debug_2"
src="https://github.com/bevyengine/bevy/assets/27962798/9d53e647-93f9-4bc7-8a20-0d9f783304d2">
---
## Changelog
* `TextAlignment` has been renamed to `JustifyText`
* `TextBundle::with_text_alignment` has been renamed to
`TextBundle::with_text_justify`
* `Text::with_alignment` has been renamed to `Text::with_justify`
* The `text_alignment` field of `TextMeasureInfo` has been renamed to
`justification`
## Migration Guide
* `TextAlignment` has been renamed to `JustifyText`
* `TextBundle::with_text_alignment` has been renamed to
`TextBundle::with_text_justify`
* `Text::with_alignment` has been renamed to `Text::with_justify`
* The `text_alignment` field of `TextMeasureInfo` has been renamed to
`justification`
# Objective
- Fixes#10806
## Solution
Replaced `new` and `index` methods for both `TableRow` and `TableId`
with `from_*` and `as_*` methods. These remove the need to perform
casting at call sites, reducing the total number of casts in the Bevy
codebase. Within these methods, an appropriate `debug_assertion` ensures
the cast will behave in an expected manner (no wrapping, etc.). I am
using a `debug_assertion` instead of an `assert` to reduce any possible
runtime overhead, however minimal. This choice is something I am open to
changing (or leaving up to another PR) if anyone has any strong
arguments for it.
---
## Changelog
- `ComponentSparseSet::sparse` stores a `TableRow` instead of a `u32`
(private change)
- Replaced `TableRow::new` and `TableRow::index` methods with
`TableRow::from_*` and `TableRow::as_*`, with `debug_assertions`
protecting any internal casting.
- Replaced `TableId::new` and `TableId::index` methods with
`TableId::from_*` and `TableId::as_*`, with `debug_assertions`
protecting any internal casting.
- All `TableId` methods are now `const`
## Migration Guide
- `TableRow::new` -> `TableRow::from_usize`
- `TableRow::index` -> `TableRow::as_usize`
- `TableId::new` -> `TableId::from_usize`
- `TableId::index` -> `TableId::as_usize`
---
## Notes
I have chosen to remove the `index` and `new` methods for the following
chain of reasoning:
- Across the codebase, `new` was called with a mixture of `u32` and
`usize` values. Likewise for `index`.
- Choosing `new` to either be `usize` or `u32` would break half of these
call-sites, requiring `as` casting at the site.
- Adding a second method `new_u32` or `new_usize` avoids the above, bu
looks visually inconsistent.
- Therefore, they should be replaced with `from_*` and `as_*` methods
instead.
Worth noting is that by updating `ComponentSparseSet`, there are now
zero instances of interacting with the inner value of `TableRow` as a
`u32`, it is exclusively used as a `usize` value (due to interactions
with methods like `len` and slice indexing). I have left the `as_u32`
and `from_u32` methods as the "proper" constructors/getters.
This removes the `From<Vec2/3>` implementations for the direction types.
It doesn't seem right to have when it only works if the vector is
nonzero and finite and produces NaN otherwise.
Added `Direction2d/3d::new` which uses `Vec2/3::try_normalize` to
guarantee it returns either a valid direction or `None`.
This should make it impossible to create an invalid direction, which I
think was the intention with these types.
Bevy builds on a ton of different libraries. Choosing who gets "prime
real estate" in the readme is hard / political. Closing that can of
worms makes a lot of sense to me.
I have also removed the "alternatives" section for similar reasons.
There are a number of engines/libs that aren't credited there, some of
those listed no longer exist / aren't maintained, and it isn't really
our job to be the canonical source. Each alternative listed is a soft
endorsement of the project (and the reverse is true for projects we
don't include). Again, closing this can of worms feels like the right
call.
# Objective
Fixes#10291
This adds a way to easily log messages once within system which are
called every frame.
## Solution
Opted for a macro-based approach. The fact that the 'once' call is
tracked per call site makes the `log_once!()` macro very versatile and
easy-to-use. I suspect it will be very handy for all of us, but
especially beginners, to get some initial feedback from systems without
spamming up the place!
I've made the macro's return its internal `has_fired` state, for
situations in which that might be useful to know (trigger something else
alongside the log, for example).
Please let me know if I placed the macro's in the right location, and if
you would like me to do something more clever with the macro's
themselves, since its looking quite copy-pastey at the moment. I've
tried ways to replace 5 with 1 macro's, but no success yet.
One downside of this approach is: Say you wish to warn the user if a
resource is invalid. In this situation, the
`resource.is_valid()` check would still be performed every frame:
```rust
fn my_system(my_res: Res<MyResource>) {
if !my_res.is_valid() {
warn_once!("resource is invalid!");
}
}
```
If you want to prevent that, you would still need to introduce a local
boolean. I don't think this is a very big deal, as expensive checks
shouldn't be called every frame in any case.
## Changelog
Added: `trace_once!()`, `debug_once!()`, `info_once!()`, `warn_once!()`,
and `error_once!()` log macros which fire only once per call site.
# Objective
Resolves Issue #10772.
## Solution
Added the deprecated warning for QueryState::for_each_unchecked, as
noted in the comments of PR #6773.
Followed the wording in the deprecation messages for `for_each` and
`for_each_mut`
# Objective
After #6547, `Query::for_each` has been capable of automatic
vectorization on certain queries, which is seeing a notable (>50% CPU
time improvements) for iteration. However, `Query::for_each` isn't
idiomatic Rust, and lacks the flexibility of iterator combinators.
Ideally, `Query::iter` and friends should be able to achieve the same
results. However, this does seem to blocked upstream
(rust-lang/rust#104914) by Rust's loop optimizations.
## Solution
This is an intermediate solution and refactor. This moves the
`Query::for_each` implementation onto the `Iterator::fold`
implementation for `QueryIter` instead. This should result in the same
automatic vectorization optimization on all `Iterator` functions that
internally use fold, including `Iterator::for_each`, `Iterator::count`,
etc.
With this, it should close the gap between the two completely.
Internally, this PR changes `Query::for_each` to use
`query.iter().for_each(..)` instead of the duplicated implementation.
Separately, the duplicate implementations of internal iteration (i.e.
`Query::par_for_each`) now use portions of the current `Query::for_each`
implementation factored out into their own functions.
This also massively cleans up our internal fragmentation of internal
iteration options, deduplicating the iteration code used in `for_each`
and `par_iter().for_each()`.
---
## Changelog
Changed: `Query::for_each`, `Query::for_each_mut`, `Query::for_each`,
and `Query::for_each_mut` have been moved to `QueryIter`'s
`Iterator::for_each` implementation, and still retains their performance
improvements over normal iteration. These APIs are deprecated in 0.13
and will be removed in 0.14.
---------
Co-authored-by: JoJoJet <21144246+JoJoJet@users.noreply.github.com>
Co-authored-by: Alice Cecile <alice.i.cecile@gmail.com>
# Objective
Make ```DynamicSceneBuilder``` more visible to new bevy learners!
```DynamicSceneBuilder``` is likely to be the most appropriate tool to use when creating dynamic scenes in all but the simplest scenarios. However, it's not mentioned in the scene example. This PR aims to fix this.
## Solution
I've modified the comment above where the ```DynamicScene``` is created to note that ```DynamicSceneBuilder``` can also be used to create the scene. I believe this is the best approach to introduce ```DynamicSceneBuilder``` without adding additional complexity to the example.
# Objective
when loading gltfs we may want to filter the results. in particular, i
need to be able to exclude cameras.
i can do this already by modifying the gltf after load and before
spawning, but it seems like a useful general option.
## Solution
add `GltfLoaderSettings` struct with bool members:
- `load_cameras` : checked before processing camera nodes.
- `load_lights` : checked before processing light nodes
- `load_meshes` : checked before loading meshes, materials and morph
weights
Existing code will work as before. Now you also have the option to
restrict what parts of the gltf are loaded. For example, to load a gltf
but exclude the cameras, replace a call to
`asset_server.load("my.gltf")` with:
```rust
asset_server.load_with_settings(
"my.gltf",
|s: &mut GltfLoaderSettings| {
s.load_cameras = false;
}
);
```
---------
Co-authored-by: Alice Cecile <alice.i.cecile@gmail.com>
# Objective
Fixes#10444
Currently manually removing an asset prevents it from being reloaded
while there are still active handles. Doing so will result in a panic,
because the storage entry has been marked as "empty / None" but the ID
is still assumed to be active by the asset server.
Patterns like `images.remove() -> asset_server.reload()` and
`images.remove() -> images.insert()` would fail if the handle was still
alive.
## Solution
Most of the groundwork for this was already laid in Bevy Asset V2. This
is largely just a matter of splitting out `remove` into two separate
operations:
* `remove_dropped`: remove the stored asset, invalidate the internal
Assets entry (preventing future insertions with the old id), and recycle
the id
* `remove_still_alive`: remove the stored asset, but leave the entry
otherwise untouched (and dont recycle the id).
`remove_still_alive` and `insert` can be called any number of times (in
any order) for an id until `remove_dropped` has been called, which will
invalidate the id.
From a user-facing perspective, there are no API changes and this is non
breaking. The public `Assets::remove` will internally call
`remove_still_alive`. `remove_dropped` can only be called by the
internal "handle management" system.
---
## Changelog
- Fix a bug preventing `Assets::remove` from blocking future inserts for
a specific `AssetIndex`.
# Objective
Fixes https://github.com/bevyengine/bevy/issues/10786
## Solution
The bind_group_layout entries for the prepass were wrong when not all 4
prepass textures were used, as it just zipped [17, 18, 19, 20] with the
smallvec of prepass `bind_group_layout` entries that potentially didn't
contain 4 entries. (eg. if you had a depth and motion vector prepass but
no normal prepass, then depth would be correct but the entry for the
motion vector prepass would be 18 (normal prepass' spot) instead of 19).
Change the prepass `get_bind_group_layout_entries` function to return an
array of `[Option<BindGroupLayoutEntryBuilder>; 4]` and only add the
layout entry if it exists.
# Objective
avoid panics from `calculate_bounds` systems if entities are despawned
in PostUpdate.
there's a running general discussion (#10166) about command panicking.
in the meantime we may as well fix up some cases where it's clear a
failure to insert is safe.
## Solution
change `.insert(aabb)` to `.try_insert(aabb)`
# Objective
> Issue raised on
[Discord](https://discord.com/channels/691052431525675048/1002362493634629796/1179182488787103776)
Currently the following code fails due to a missing `TypePath` bound:
```rust
#[derive(Reflect)] struct Foo<T>(T);
#[derive(Reflect)] struct Bar<T>(Foo<T>);
#[derive(Reflect)] struct Baz<T>(Bar<Foo<T>>);
```
## Solution
Add `TypePath` to the per-field bounds instead of _just_ the generic
type parameter bounds.
### Related Work
It should be noted that #9046 would help make these kinds of issues
easier to work around and/or avoid entirely.
---
## Changelog
- Fixes missing `TypePath` requirement when deriving `Reflect` on nested
generics
# Objective
Fixes#10401
## Solution
* Allow sources to register specific processed/unprocessed watch
warnings.
* Specify per-platform watch warnings. This removes the need to cover
all platform cases in one warning message.
* Only register watch warnings for the _processed_ embedded source, as
warning about watching unprocessed embedded isn't helpful.
---
## Changelog
- Asset sources can now register specific watch warnings.
# Objective
- Shorten paths by removing unnecessary prefixes
## Solution
- Remove the prefixes from many paths which do not need them. Finding
the paths was done automatically using built-in refactoring tools in
Jetbrains RustRover.
# Objective
These type are unavailable to editors and scripting interfaces making
use of reflection.
## Solution
`#[derive(Reflect)]` and call `.register_type` during plugin
initialization.
---
## Changelog
### Added
- Implement `Reflect` for audio-related types, and register them.
# Objective
- Materials should be a more frequent rebind then meshes (due to being
able to use a single vertex buffer, such as in #10164) and therefore
should be in a higher bind group.
---
## Changelog
- For 2d and 3d mesh/material setups (but not UI materials, or other
rendering setups such as gizmos, sprites, or text), mesh data is now in
bind group 1, and material data is now in bind group 2, which is swapped
from how they were before.
## Migration Guide
- Custom 2d and 3d mesh/material shaders should now use bind group 2
`@group(2) @binding(x)` for their bound resources, instead of bind group
1.
- Many internal pieces of rendering code have changed so that mesh data
is now in bind group 1, and material data is now in bind group 2.
Semi-custom rendering setups (that don't use the Material or Material2d
APIs) should adapt to these changes.
# Objective
Keep essentially the same structure of `EntityHasher` from #9903, but
rephrase the multiplication slightly to save an instruction.
cc @superdump
Discord thread:
https://discord.com/channels/691052431525675048/1172033156845674507/1174969772522356756
## Solution
Today, the hash is
```rust
self.hash = i | (i.wrapping_mul(FRAC_U64MAX_PI) << 32);
```
with `i` being `(generation << 32) | index`.
Expanding things out, we get
```rust
i | ( (i * CONST) << 32 )
= (generation << 32) | index | ((((generation << 32) | index) * CONST) << 32)
= (generation << 32) | index | ((index * CONST) << 32) // because the generation overflowed
= (index * CONST | generation) << 32 | index
```
What if we do the same thing, but with `+` instead of `|`? That's almost
the same thing, except that it has carries, which are actually often
better in a hash function anyway, since it doesn't saturate. (`|` can be
dangerous, since once something becomes `-1` it'll stay that, and
there's no mixing available.)
```rust
(index * CONST + generation) << 32 + index
= (CONST << 32 + 1) * index + generation << 32
= (CONST << 32 + 1) * index + (WHATEVER << 32 + generation) << 32 // because the extra overflows and thus can be anything
= (CONST << 32 + 1) * index + ((CONST * generation) << 32 + generation) << 32 // pick "whatever" to be something convenient
= (CONST << 32 + 1) * index + ((CONST << 32 + 1) * generation) << 32
= (CONST << 32 + 1) * index +((CONST << 32 + 1) * (generation << 32)
= (CONST << 32 + 1) * (index + generation << 32)
= (CONST << 32 + 1) * (generation << 32 | index)
= (CONST << 32 + 1) * i
```
So we can do essentially the same thing using a single multiplication
instead of doing multiply-shift-or.
LLVM was already smart enough to merge the shifting into a
multiplication, but this saves the extra `or`:
<https://rust.godbolt.org/z/MEvbz4eo4>
It's a very small change, and often will disappear in load latency
anyway, but it's a couple percent faster in lookups:
(There was more of an improvement here before #10558, but with `to_bits`
being a single `qword` load now, keeping things mostly as it is turned
out to be better than the bigger changes I'd tried in #10605.)
---
## Changelog
(Probably skip it)
## Migration Guide
(none needed)
# Objective
the pbr prepass vertex shader currently only sets
`VertexOutput::world_position` when deferred or motion prepasses are
enabled.
the field is always in the vertex output so is otherwise undetermined,
and the calculation is very cheap.
## Solution
always set the world position in the pbr prepass vert shader.
