# Objective
The `post_processing` example is currently broken when run with webgl2.
```
cargo run --example post_processing --target=wasm32-unknown-unknown
```
```
wasm.js:387 panicked at 'wgpu error: Validation Error
Caused by:
In Device::create_render_pipeline
note: label = `post_process_pipeline`
In the provided shader, the type given for group 0 binding 2 has a size of 4. As the device does not support `DownlevelFlags::BUFFER_BINDINGS_NOT_16_BYTE_ALIGNED`, the type must have a size that is a multiple of 16 bytes.
```
I bisected the breakage to c7eaedd6a1.
## Solution
Add padding when using webgl2
# Objective
This doc comment for the `set` method of `ContentSize`:
```
Set a `Measure` for this function
```
doesn't seem to make sense, `ContentSize` is not a function.
# Solution
Replace it.
# Objective
When an `AudioSink` is removed from an entity, the audio player will
automatically start any `AudioSource` still attached, which normally is
the one used to start playback in the first place.
## Solution
Long story short, the default behavior is restarting the audio, and this
commit documents that.
---
## Changelog
Fixed documentation on `AudioSink` to clarify removal behavior.
# Objective
- Some examples crash in CI because of needing too many resources for
the windows runner
- Some examples have random results making it hard to compare
screenshots
## Solution
- `bloom_3d`: reduce the number of spheres
- `pbr`: use simpler spheres and reuse the mesh
- `tonemapping`: use simpler spheres and reuse the mesh
- `shadow_biases`: reduce the number of spheres
- `spotlight`: use a seeded rng, move more cubes in view while reducing
the total number of cubes, and reuse meshes and materials
- `external_source_external_thread`, `iter_combinations`,
`parallel_query`: use a seeded rng
Examples of errors encountered:
```
Caused by:
In Device::create_bind_group
note: label = `bloom_upsampling_bind_group`
Not enough memory left
```
```
Caused by:
In Queue::write_buffer
Parent device is lost
```
```
ERROR wgpu_core::device::life: Mapping failed Device(Lost)
```
# Objective
shader defs associated with a shader via `load_internal_asset!` or
`Shader::from_xxx_with_defs` were being accidentally ignored for
top-level shaders.
## Solution
include the defs for top level shaders.
# Objective
- Fixes#9114
## Solution
Inside `ScheduleGraph::build_schedule()` the variable `node_count =
self.systems.len() + self.system_sets.len()` is used to calculate the
indices for the `reachable` bitset derived from `self.hierarchy.graph`.
However, the number of nodes inside `self.hierarchy.graph` does not
always correspond to `self.systems.len() + self.system_sets.len()` when
`ambiguous_with` is used, because an ambiguous set is added to
`system_sets` (because we need an `NodeId` for the ambiguity graph)
without adding a node to `self.hierarchy`.
In this PR, we rename `node_count` to the more descriptive name
`hg_node_count` and set it to `self.hierarchy.graph.node_count()`.
---------
Co-authored-by: James Liu <contact@jamessliu.com>
# Objective
Fixes#9113
## Solution
disable `multi-threaded` default feature
## Migration Guide
The `multi-threaded` feature in `bevy_ecs` and `bevy_tasks` is no longer
enabled by default. However, this remains a default feature for the
umbrella `bevy` crate. If you depend on `bevy_ecs` or `bevy_tasks`
directly, you should consider enabling this to allow systems to run in
parallel.
Fixes#5856. Fixes#8080. Fixes#9040.
# Objective
We need to limit the update rate of games whose windows are not visible
(minimized or completely occluded). Compositors typically ignore the
VSync settings of windows that aren't visible. That, combined with the
lack of rendering work, results in a scenario where an app becomes
completely CPU-bound and starts updating without limit.
There are currently three update modes.
- `Continuous` updates an app as often as possible.
- `Reactive` updates when new window or device events have appeared, a
timer expires, or a redraw is requested.
- `ReactiveLowPower` is the same as `Reactive` except it ignores device
events (e.g. general mouse movement).
The problem is that the default "game" settings are set to `Contiuous`
even when no windows are visible.
### More Context
- https://github.com/libsdl-org/SDL/issues/1871
- https://github.com/glfw/glfw/issues/680
- https://github.com/godotengine/godot/issues/19741
- https://github.com/godotengine/godot/issues/64708
## Solution
Change the default "unfocused" `UpdateMode` for games to
`ReactiveLowPower` just like desktop apps. This way, even when the
window is occluded, the app still updates at a sensible rate or if
something about the window changes. I chose 20Hz arbitrarily.
# Objective
The `lifetimeless` module has been a source of confusion for bevy users
for a while now.
## Solution
Add a couple paragraph explaining that, yes, you can use one of the type
alias safely, without ever leaking any memory.
Redo of #7590 since I messed up my branch.
# Objective
- Revise docs.
- Refactor event loop code a little bit to make it easier to follow.
## Solution
- Do the above.
---
### Migration Guide
- `UpdateMode::Reactive { max_wait: .. }` -> `UpdateMode::Reactive {
wait: .. }`
- `UpdateMode::ReactiveLowPower { max_wait: .. }` ->
`UpdateMode::ReactiveLowPower { wait: .. }`
---------
Co-authored-by: Sélène Amanita <134181069+Selene-Amanita@users.noreply.github.com>
# Objective
Remove the `With<Parent>` query filter from the `parent_node_query`
parameter of the `bevy_ui::render::extract_uinode_borders` function.
This is a bug, the query is only used to retrieve the size of the
current node's parent. We don't care if that parent node has a `Parent`
or not.
---------
Co-authored-by: François <mockersf@gmail.com>
# Objective
- Fixes#9138
## Solution
- Calling `Camera2dBundle::default()` will now result in a
`Camera2dBundle` with `Vec3::ZERO` transform, `far` value of `1000.` and
`near` value of `-1000.`.
- This will enable the rendering of 2d entities in negative z space by
default.
- I did not modify `new_with_far` as moving the camera to `Vec3::ZERO`
in that function will cause entities in the positive z space to become
hidden without further changes. And the further changes cannot be
applied without it being a breaking change.
# Objective
Glam 0.24 added new glam types (```I64Vec``` and ```U64Vec```). However
these are not reflectable unlike the other glam types
## Solution
Implement reflect for these new types
---
## Changelog
Implements reflect with the impl_reflect_struct macro on ```I64Vec2```,
```I64Vec3```, ```I64Vec4```, ```U64Vec2```, ```U64Vec3```, and
```U64Vec4``` types
# Objective
- Repeat in `Gizmos` that they are drawned in immediate mode, which is
said at the module level but not here, and detail what it means.
- Clarify for every method of `Gizmos` that they should be called for
every frame.
- Clarify which methods belong to 3D or 2D space (kinda obvious for 2D
but still)
The first time I used gizmos I didn't understand how they work and was
confused as to why nothing showed up.
---------
Co-authored-by: François <mockersf@gmail.com>
Co-authored-by: SpecificProtagonist <vincentjunge@posteo.net>
Co-authored-by: Alice Cecile <alice.i.cecile@gmail.com>
# Objective
- In bevy_polyline, we discovered an issue that happens when line width
is smaller than 1.0 and using perspective. It would sometimes end up
negative or NaN. I'm not entirely sure _why_ it happens.
## Solution
- Make sure the width doesn't go below 0 before multiplying it with the
alpha
# Notes
Here's a link to the bevy_polyline issue
https://github.com/ForesightMiningSoftwareCorporation/bevy_polyline/issues/46
I'm not sure if the solution is correct but it solved the issue in my
testing.
Co-authored-by: François <mockersf@gmail.com>
# Objective
This PR continues https://github.com/bevyengine/bevy/pull/8885
It aims to improve the `Mesh` documentation in the following ways:
- Put everything at the "top level" instead of the "impl".
- Explain better what is a Mesh, how it can be created, and that it can
be edited.
- Explain it can be used with a `Material`, and mention
`StandardMaterial`, `PbrBundle`, `ColorMaterial`, and
`ColorMesh2dBundle` since those cover most cases
- Mention the glTF/Bevy vocabulary discrepancy for "Mesh"
- Add an image for the example
- Various nitpicky modifications
## Note
- The image I added is 90.3ko which I think is small enough?
- Since rustdoc doesn't allow cross-reference not in dependencies of a
subcrate [yet](https://github.com/rust-lang/rust/issues/74481), I have a
lot of backtick references that are not links :(
- Since rustdoc doesn't allow linking to code in the crate (?) I put
link to github directly.
- Since rustdoc doesn't allow embed images in doc
[yet](https://github.com/rust-lang/rust/issues/32104), maybe
[soon](https://github.com/rust-lang/rfcs/pull/3397), I had to put only a
link to the image. I don't think it's worth adding
[embed_doc_image](https://docs.rs/embed-doc-image/latest/embed_doc_image/)
as a dependency for this.
# Objective
- Fix shader_material_glsl example
## Solution
- Expose the `PER_OBJECT_BUFFER_BATCH_SIZE` shader def through the
default `MeshPipeline` specialization.
- Make use of it in the `custom_material.vert` shader to access the mesh
binding.
---
## Changelog
- Added: Exposed the `PER_OBJECT_BUFFER_BATCH_SIZE` shader def through
the default `MeshPipeline` specialization to use in custom shaders not
using bevy_pbr::mesh_bindings that still want to use the mesh binding in
some way.
# Objective
- The `path` module was getting fairly large.
- The code in `AccessRef::read_element` and mut equivalent was very
complex and difficult to understand.
- The `ReflectPathError` had a lot of variants, and was difficult to
read.
## Solution
- Split the file in two, `access` now has its own module
- Rewrite the `read_element` methods, they were ~200 lines long, they
are now ~70 lines long — I didn't change any of the logic. It's really
just the same code, but error handling is separated.
- Split the `ReflectPathError` error
- Merge `AccessRef` and `Access`
- A few other changes that aim to reduce code complexity
### Fully detailed change list
- `Display` impl of `ParsedPath` now includes prefix dots — this allows
simplifying its implementation, and IMO `.path.to.field` is a better way
to express a "path" than `path.to.field` which could suggest we are
reading the `to` field of a variable named `path`
- Add a test to check that dot prefixes and other are correctly parsed —
Until now, no test contained a prefixing dot
- Merge `Access` and `AccessRef`, using a `Cow<'a, str>`. Generated code
seems to agree with this decision (`ParsedPath::parse` sheds 5% of
instructions)
- Remove `Access::as_ref` since there is no such thing as an `AccessRef`
anymore.
- Rename `AccessRef::to_owned` into `AccessRef::into_owned()` since it
takes ownership of `self` now.
- Add a `parse_static` that doesn't allocate new strings for named
fields!
- Add a section about path reflection in the `bevy_reflect` crate root
doc — I saw a few people that weren't aware of path reflection, so I
thought it was pertinent to add it to the root doc
- a lot of nits
- rename `index` to `offset` when it refers to offset in the path string
— There is no more confusion with the other kind of indices in this
context, also it's a common naming convention for parsing.
- Make a dedicated enum for parsing errors
- rename the `read_element` methods to `element` — shorter, but also
`read_element_mut` was a fairly poor name
- The error values now not only contain the expected type but also the
actual type.
- Remove lifetimes that could be inferred from the `GetPath` trait
methods.
---
## Change log
- Added the `ParsedPath::parse_static` method, avoids allocating when
parsing `&'static str`.
## Migration Guide
If you were matching on the `Err(ReflectPathError)` value returned by
`GetPath` and `ParsedPath` methods, now only the parse-related errors
and the offset are publicly accessible. You can always use the
`fmt::Display` to get a clear error message, but if you need
programmatic access to the error types, please open an issue.
---------
Co-authored-by: Alice Cecile <alice.i.cecile@gmail.com>
Co-authored-by: Gino Valente <49806985+MrGVSV@users.noreply.github.com>
- attempt to clarify with better docstrings the default behaviour of
WindowPlugin and the component type it accepts.
# Objective
- I'm new to Rust and Bevy, I got a bit confused about how to customise
some window parameters (title, height, width etc) and while the docs do
show the struct code for that field `Option<Window>` I was a bit of a
doofus and skipped that to read the docstring entry for `primary_window`
and then the `Window` component directly which aren't linked together.
This is a minor improvement which I think will help in-case others, like
me, have a doofus moment.
---------
Co-authored-by: Sélène Amanita <134181069+Selene-Amanita@users.noreply.github.com>
# Objective
The documentation for the `print_when_completed` system stated that this
system would tick the `Timer` component on every entity in the scene.
This was incorrect as this system only ticks the `Timer` on entities
with the `PrintOnCompletionTimer` component.
## Solution
We suggest a modification to the documentation of this system to make it
more clear.
# Objective
This PR updates the name of the enum variant used in the docs for
`OrthographicProjection`.
## Solution
- Change the outdated 'WindowScale` to `WindowSize`.
# Objective
- Reduce the number of rebindings to enable batching of draw commands
## Solution
- Use the new `GpuArrayBuffer` for `MeshUniform` data to store all
`MeshUniform` data in arrays within fewer bindings
- Sort opaque/alpha mask prepass, opaque/alpha mask main, and shadow
phases also by the batch per-object data binding dynamic offset to
improve performance on WebGL2.
---
## Changelog
- Changed: Per-object `MeshUniform` data is now managed by
`GpuArrayBuffer` as arrays in buffers that need to be indexed into.
## Migration Guide
Accessing the `model` member of an individual mesh object's shader
`Mesh` struct the old way where each `MeshUniform` was stored at its own
dynamic offset:
```rust
struct Vertex {
@location(0) position: vec3<f32>,
};
fn vertex(vertex: Vertex) -> VertexOutput {
var out: VertexOutput;
out.clip_position = mesh_position_local_to_clip(
mesh.model,
vec4<f32>(vertex.position, 1.0)
);
return out;
}
```
The new way where one needs to index into the array of `Mesh`es for the
batch:
```rust
struct Vertex {
@builtin(instance_index) instance_index: u32,
@location(0) position: vec3<f32>,
};
fn vertex(vertex: Vertex) -> VertexOutput {
var out: VertexOutput;
out.clip_position = mesh_position_local_to_clip(
mesh[vertex.instance_index].model,
vec4<f32>(vertex.position, 1.0)
);
return out;
}
```
Note that using the instance_index is the default way to pass the
per-object index into the shader, but if you wish to do custom rendering
approaches you can pass it in however you like.
---------
Co-authored-by: robtfm <50659922+robtfm@users.noreply.github.com>
Co-authored-by: Elabajaba <Elabajaba@users.noreply.github.com>
# Objective
My attempt at implementing #7515
## Solution
Added struct `Pitch` and implemented on it `Source` trait.
## Changelog
### Added
- File pitch.rs to bevy_audio crate
- Struct `Pitch` and type aliases for `AudioSourceBundle<Pitch>` and
`SpatialAudioSourceBundle<Pitch>`
- New example showing how to use `PitchBundle`
### Changed
- `AudioPlugin` now adds system for `Pitch` audio
---------
Co-authored-by: Alice Cecile <alice.i.cecile@gmail.com>
The previous formatting didn't render as you'd expect, with 'For CSS
Grid containers' getting adopted by the prior bullet point. Rather than
fixing that directly I opted for a slight reformatting for consistency
with other fields, notably left/right/top/bottom.
# Objective
Fixes#9298 - Default window title leaks "bevy" context
## Solution
I just replaced the literal string "Bevy App" with "App" in Window's
Default trait implementation.
# Objective
Cloning a `WorldQuery` type's "fetch" struct was made unsafe in #5593,
by adding the `unsafe fn clone_fetch` to `WorldQuery`. However, as that
method's documentation explains, it is not the right place to put the
safety invariant:
> While calling this method on its own cannot cause UB it is marked
`unsafe` as the caller must ensure that the returned value is not used
in any way that would cause two `QueryItem<Self>` for the same
`archetype_index` or `table_row` to be alive at the same time.
You can clone a fetch struct all you want and it will never cause
undefined behavior -- in order for something to go wrong, you need to
improperly call `WorldQuery::fetch` with it (which is marked unsafe).
Additionally, making it unsafe to clone a fetch struct does not even
prevent undefined behavior, since there are other ways to incorrectly
use a fetch struct. For example, you could just call fetch more than
once for the same entity, which is not currently forbidden by any
documented invariants.
## Solution
Document a safety invariant on `WorldQuery::fetch` that requires the
caller to not create aliased `WorldQueryItem`s for mutable types. Remove
the `clone_fetch` function, and add the bound `Fetch: Clone` instead.
---
## Changelog
- Removed the associated function `WorldQuery::clone_fetch`, and added a
`Clone` bound to `WorldQuery::Fetch`.
## Migration Guide
### `fetch` invariants
The function `WorldQuery::fetch` has had the following safety invariant
added:
> If this type does not implement `ReadOnlyWorldQuery`, then the caller
must ensure that it is impossible for more than one `Self::Item` to
exist for the same entity at any given time.
This invariant was always required for soundness, but was previously
undocumented. If you called this function manually anywhere, you should
check to make sure that this invariant is not violated.
### Removed `clone_fetch`
The function `WorldQuery::clone_fetch` has been removed. The associated
type `WorldQuery::Fetch` now has the bound `Clone`.
Before:
```rust
struct MyFetch<'w> { ... }
unsafe impl WorldQuery for MyQuery {
...
type Fetch<'w> = MyFetch<'w>
unsafe fn clone_fetch<'w>(fetch: &Self::Fetch<'w>) -> Self::Fetch<'w> {
MyFetch {
field1: fetch.field1,
field2: fetch.field2.clone(),
...
}
}
}
```
After:
```rust
#[derive(Clone)]
struct MyFetch<'w> { ... }
unsafe impl WorldQuery for MyQuery {
...
type Fetch<'w> = MyFetch<'w>;
}
```
---------
Co-authored-by: Alice Cecile <alice.i.cecile@gmail.com>
# Objective
Fixes https://github.com/bevyengine/bevy/issues/9234
re-breaks: The issues that were linked in #9169
## Solution
Revert the PR that broke tonemapping/postprocessing/etc.
Any passes that are post msaa resolve need to use the main textures, not
the msaa texture.
## Changelog
Idk what to put here since it's a revert.
# Objective
- CI job `check-bans` fails often for unrelated reasons to a PR. Reduce
those failures
## Solution
- Currently, the job only runs if a `Cargo.toml` file changed. This PR
would run the job only if the output of `cargo tree --depth 3` is
different in a PR from main. the job would still always run on main
# Objective
Fixes#9097
## Solution
Reorder the `ExtractSchedule` so that the `extract_text_uinodes` and
`extract_uinode_borders` systems are run after `extract_atlas_uinodes`.
## Changelog
`bevy_ui::render`:
* Added the `ExtractAtlasNode` variant to `RenderUiSystem`.
* Changed `ExtractSchedule` so that `extract_uinode_borders` and
`extract_text_uinodes` run after `extract_atlas_uinodes`.
This is not used directly within the rendering code.
# Objective
- Remove extraneous dependency on `wgpu-hal` as it is not used.
## Solution
- The dependency has been removed and should have no externally visible
impact.
# Objective
The `QueryParIter::for_each_mut` function is required when doing
parallel iteration with mutable queries.
This results in an unfortunate stutter:
`query.par_iter_mut().par_for_each_mut()` ('mut' is repeated).
## Solution
- Make `for_each` compatible with mutable queries, and deprecate
`for_each_mut`. In order to prevent `for_each` from being called
multiple times in parallel, we take ownership of the QueryParIter.
---
## Changelog
- `QueryParIter::for_each` is now compatible with mutable queries.
`for_each_mut` has been deprecated as it is now redundant.
## Migration Guide
The method `QueryParIter::for_each_mut` has been deprecated and is no
longer functional. Use `for_each` instead, which now supports mutable
queries.
```rust
// Before:
query.par_iter_mut().for_each_mut(|x| ...);
// After:
query.par_iter_mut().for_each(|x| ...);
```
The method `QueryParIter::for_each` now takes ownership of the
`QueryParIter`, rather than taking a shared reference.
```rust
// Before:
let par_iter = my_query.par_iter().batching_strategy(my_batching_strategy);
par_iter.for_each(|x| {
// ...Do stuff with x...
par_iter.for_each(|y| {
// ...Do nested stuff with y...
});
});
// After:
my_query.par_iter().batching_strategy(my_batching_strategy).for_each(|x| {
// ...Do stuff with x...
my_query.par_iter().batching_strategy(my_batching_strategy).for_each(|y| {
// ...Do nested stuff with y...
});
});
```
# Objective
Fixes#9121
Context:
- `ImageTextureLoader` depends on `RenderDevice` to work out which
compressed image formats it can support
- `RenderDevice` is initialised by `RenderPlugin`
- https://github.com/bevyengine/bevy/pull/8336 made `RenderPlugin`
initialisation async
- This caused `RenderDevice` to be missing at the time of
`ImageTextureLoader` initialisation, which in turn meant UASTC encoded
ktx2 textures were being converted to unsupported formats, and thus
caused panics
## Solution
- Delay `ImageTextureLoader` initialisation
---
## Changelog
- Moved `ImageTextureLoader` initialisation from `ImagePlugin::build()`
to `ImagePlugin::finish()`
- Default to `CompressedImageFormats::NONE` if `RenderDevice` resource
is missing
---------
Co-authored-by: 66OJ66 <hi0obxud@anonaddy.me>
### **Adopted #6430**
# Objective
`MutUntyped` is the untyped variant of `Mut<T>` that stores a `PtrMut`
instead of a `&mut T`. Working with a `MutUntyped` is a bit annoying,
because as soon you want to use the ptr e.g. as a `&mut dyn Reflect` you
cannot use a type like `Mut<dyn Reflect>` but instead need to carry
around a `&mut dyn Reflect` and a `impl FnMut()` to mark the value as
changed.
## Solution
* Provide a method `map_unchanged` to turn a `MutUntyped` into a
`Mut<T>` by mapping the `PtrMut<'a>` to a `&'a mut T`
This can be used like this:
```rust
// SAFETY: ptr is of type `u8`
let val: Mut<u8> = mut_untyped.map_unchanged(|ptr| unsafe { ptr.deref_mut::<u8>() });
// SAFETY: from the context it is known that `ReflectFromPtr` was made for the type of the `MutUntyped`
let val: Mut<dyn Reflect> = mut_untyped.map_unchanged(|ptr| unsafe { reflect_from_ptr.as_reflect_ptr_mut(ptr) });
```
Note that nothing prevents you from doing
```rust
mut_untyped.map_unchanged(|ptr| &mut ());
```
or using any other mutable reference you can get, but IMO that is fine
since that will only result in a `Mut` that will dereference to that
value and mark the original value as changed. The lifetimes here prevent
anything bad from happening.
## Alternatives
1. Make `Ticks` public and provide a method to get construct a `Mut`
from `Ticks` and `&mut T`. More powerful and more easy to misuse.
2. Do nothing. People can still do everything they want, but they need
to pass (`&mut dyn Reflect, impl FnMut() + '_)` around instead of
`Mut<dyn Reflect>`
## Changelog
- add `MutUntyped::map_unchanged` to turn a `MutUntyped` into its typed
counterpart
---------
Co-authored-by: Jakob Hellermann <jakob.hellermann@protonmail.com>
Co-authored-by: JoJoJet <21144246+JoJoJet@users.noreply.github.com>
# Objective
Fix#8936.
## Solution
Stop using `unwrap` in the core pipelined rendering logic flow.
Separately also scoped the `sub app` span to just running the render app
instead of including the blocking send.
Current unknowns: should we use `std::panic::catch_unwind` around
running the render app? Other engine threads use it defensively, but
we're letting it bubble up here, and a user-created panic could cause a
deadlock if it kills the thread.
---
## Changelog
Fixed: Pipelined rendering should no longer have spurious panics upon
app exit.
# Objective
Implements #9082 but with an option to toggle minimize and close buttons
too.
## Solution
- Added an `enabled_buttons` member to the `Window` struct through which
users can enable or disable specific window control buttons.
---
## Changelog
- Added an `enabled_buttons` member to the `Window` struct through which
users can enable or disable specific window control buttons.
- Added a new system to the `window_settings` example which demonstrates
the toggling functionality.
---
## Migration guide
- Added an `enabled_buttons` member to the `Window` struct through which
users can enable or disable specific window control buttons.
# Objective
Currently the panic message if a duplicate plugin is added isn't really
helpful or at least can be made more useful if it includes the location
where the plugin was added a second time.
## Solution
Add `track_caller` to `add_plugins` and it's called dependencies.
# Objective
This attempts to make the new IRect and URect structs in bevy_math more
similar to the existing Rect struct.
## Solution
Add reflect implementations for IRect and URect, since one already
exists for Rect.
# Objective
- #8960 isn't optimal for very distinct AABB colors, it can be improved
## Solution
We want a function that maps sequential values (entities concurrently
living in a scene _usually_ have ids that are sequential) into very
different colors (the hue component of the color, to be specific)
What we are looking for is a [so-called "low discrepancy"
sequence](https://en.wikipedia.org/wiki/Low-discrepancy_sequence). ie: a
function `f` such as for integers in a given range (eg: 101, 102, 103…),
`f(i)` returns a rational number in the [0..1] range, such as `|f(i) -
f(i±1)| ≈ 0.5` (maximum difference of images for neighboring preimages)
AHash is a good random hasher, but it has relatively high discrepancy,
so we need something else.
Known good low discrepancy sequences are:
#### The [Van Der Corput
sequence](https://en.wikipedia.org/wiki/Van_der_Corput_sequence)
<details><summary>Rust implementation</summary>
```rust
fn van_der_corput(bits: u64) -> f32 {
let leading_zeros = if bits == 0 { 0 } else { bits.leading_zeros() };
let nominator = bits.reverse_bits() >> leading_zeros;
let denominator = bits.next_power_of_two();
nominator as f32 / denominator as f32
}
```
</details>
#### The [Gold Kronecker
sequence](https://extremelearning.com.au/unreasonable-effectiveness-of-quasirandom-sequences/)
<details><summary>Rust implementation</summary>
Note that the implementation suggested in the linked post assumes
floats, we have integers
```rust
fn gold_kronecker(bits: u64) -> f32 {
const U64_MAX_F: f32 = u64::MAX as f32;
// (u64::MAX / Φ) rounded down
const FRAC_U64MAX_GOLDEN_RATIO: u64 = 11400714819323198485;
bits.wrapping_mul(FRAC_U64MAX_GOLDEN_RATIO) as f32 / U64_MAX_F
}
```
</details>
### Comparison of the sequences
So they are both pretty good. Both only have a single (!) division and
two `u32 as f32` conversions.
- Kronecker is resilient to regular sequence (eg: 100, 102, 104, 106)
while this kills Van Der Corput (consider that potentially one entity
out of two spawned might be a mesh)
I made a small app to compare the two sequences, available at:
https://gist.github.com/nicopap/5dd9bd6700c6a9a9cf90c9199941883e
At the top, we have Van Der Corput, at the bottom we have the Gold
Kronecker. In the video, we spawn a vertical line at the position on
screen where the x coordinate is the image of the sequence. The
preimages are 1,2,3,4,… The ideal algorithm would always have the
largest possible gap between each line (imagine the screen x coordinate
as the color hue):
https://github.com/bevyengine/bevy/assets/26321040/349aa8f8-f669-43ba-9842-f9a46945e25c
Here, we repeat the experiment, but with with `entity.to_bits()` instead
of a sequence:
https://github.com/bevyengine/bevy/assets/26321040/516cea27-7135-4daa-a4e7-edfd1781d119
Notice how Van Der Corput tend to bunch the lines on a single side of
the screen. This is because we always skip odd-numbered entities.
Gold Kronecker seems always worse than Van Der Corput, but it is
resilient to finicky stuff like entity indices being multiples of a
number rather than purely sequential, so I prefer it over Van Der
Corput, since we can't really predict how distributed the entity indices
will be.
### Chosen implementation
You'll notice this PR's implementation is not the Golden ratio-based
Kronecker sequence as described in
[tueoqs](https://extremelearning.com.au/unreasonable-effectiveness-of-quasirandom-sequences/).
Why?
tueoqs R function multiplies a rational/float and takes the fractional
part of the result `(x/Φ) % 1`. We start with an integer `u32`. So
instead of converting into float and dividing by Φ (mod 1) we directly
divide by Φ as integer (mod 2³²) both operations are equivalent, the
integer division (which is actually a multiplication by `u32::MAX / Φ`)
is probably faster.
## Acknowledgements
- `inspi` on discord linked me to
https://extremelearning.com.au/unreasonable-effectiveness-of-quasirandom-sequences/
and the wikipedia article.
- [this blog
post](https://probablydance.com/2018/06/16/fibonacci-hashing-the-optimization-that-the-world-forgot-or-a-better-alternative-to-integer-modulo/)
for the idea of multiplying the `u32` rather than the `f32`.
- `nakedible` for suggesting the `index()` over `to_bits()` which
considerably reduces generated code (goes from 50 to 11 instructions)
