Add bindless support for 2D materials. by pcwalton · Pull Request #24833 · bevyengine/bevy

pcwalton · 2026-07-01T22:23:01Z

At the moment, Bevy must issue a separate drawcall and bind group switch for each 2D object with a different texture, incurring significant wgpu and driver overhead. To address this problem, this commit ports the bindless support that the 3D PBR pipeline has long enjoyed to the 2D pipeline as well. Bindless allows Bevy to batch 2D meshes into single drawcalls as long as they share the same mesh, even if their materials (including textures) differ.

To implement bindless, this patch moves the efficient material bind group allocator from bevy_pbr to bevy_render. Like the 3D PBR pipeline, the code in this PR checks to see whether bindless is supported on the current platform and automatically falls back to non-bindless if it isn't.

Because the logic needed to support bindless is tightly intertwined with the logic needed to prepare pipelines, this PR also overhauls specialize_material2d_meshes. That function is now very similar to the one in the 3D pipeline. At some point, it'd probably be best to merge the core logic of this function for the 2D and 3D pipelines together. However, I opted to leave that to a follow-up for two reasons: (1) it's easier to verify in review that the 2D and 3D functions are the same than to verify that any added layers of abstraction are correct; (2) it might (or might not) be best to wait until multi-draw indirect is added to the 2D pipeline to do the refactoring.

Per review request, I've left the migration of ColorMaterial and SpriteMaterial over to bindless to a follow-up. With this follow-up (not in this PR), on the bevymark example with --waves 1 --per-wave 10000 --vary-per-instance --material-texture-count 1000, I observed:

With the dedicated sprite rendering path, which doesn't use bindless, I get 21.5 ms/frame, or 47 FPS.
With the SpriteMaterial path, with the follow-up patch that adds bindless support to that material, I get 17.4 ms/frame, or 67 FPS.
With the SpriteMaterial path on main, I get 29.7 ms/frame, or 38.2 FPS.

The follow-up patch is therefore a 1.71× speedup on that workload for SpriteMaterial and a 1.24× speedup over the dedicated sprite rendering path.

Note that, in that follow-up patch, order to see the changes in action, I had to change bevymark to avoid using the same seed for all its parallel RNGs. Without that change, the fact that each RNG uses the same seed means that each texture effectively gets its own Z layer, causing the benchmark to degenerate into the optimal scenario for batching in the non-bindless case. The follow-up PR scrambles the RNG seeds to avoid this pathological, unrealistic behavior.

Zeophlite · 2026-07-01T22:52:37Z

+/// The material bind group allocator is infrastructure for bindless resources.
+/// It packs multiple materials into a small number of bind groups, allowing
+/// Bevy to render large parts of the scene with a small number of drawcalls.
+pub struct MaterialBindGroupsPlugin;


Could you split this PR - one to move material bind groups from pbr to render

The other to update colour and sprite materials to use it

OK, I ripped the ColorMaterial and SpriteMaterial update out of this PR. It’ll be a follow-up.

Cheers, it just makes it easier to follow

github-actions · 2026-07-01T22:58:02Z

The generated examples/README.md is out of sync with the example metadata in Cargo.toml or the example readme template. Please run cargo run -p build-templated-pages -- update examples to update it, and commit the file change.

At the moment, Bevy must issue a separate drawcall and bind group switch for each 2D object with a different texture, incurring significant `wgpu` and driver overhead. To address this problem, this commit ports the bindless support that the 3D PBR pipeline has long enjoyed to the 2D pipeline as well. Bindless allows Bevy to batch 2D meshes into single drawcalls as long as they share the same mesh, even if their materials (including textures) differ. To implement bindless, this patch moves the efficient material bind group allocator from `bevy_pbr` to `bevy_render`. Like the 3D PBR pipeline, the code in this PR checks to see whether bindless is supported on the current platform and automatically falls back to non-bindless if it isn't. Because the logic needed to support bindless is tightly intertwined with the logic needed to prepare pipelines, this PR also overhauls `specialize_material2d_meshes`. That function is now very similar to the one in the 3D pipeline. At some point, it'd probably be best to merge the core logic of this function for the 2D and 3D pipelines together. However, I opted to leave that to a follow-up for two reasons: (1) it's easier to verify in review that the 2D and 3D functions are the same than to verify that any added layers of abstraction are correct; (2) it might (or might not) be best to wait until multi-draw indirect is added to the 2D pipeline to do the refactoring. Per review request, I've left the migration of `ColorMaterial` and `SpriteMaterial` over to bindless to a follow-up. With this follow-up (*not in this PR*), on the `bevymark` example with `--waves 1 --per-wave 10000 --vary-per-instance --material-texture-count 1000`, I observed: * With the dedicated sprite rendering path, which doesn't use bindless, I get 21.5 ms/frame, or 47 FPS. * With the `SpriteMaterial` path, with the follow-up patch that adds bindless support to that material, I get 17.4 ms/frame, or 67 FPS. * With the `SpriteMaterial` path on `main`, I get 29.7 ms/frame, or 38.2 FPS. The follow-up patch is therefore a 1.71× speedup on that workload for `SpriteMaterial` and a 1.24× speedup over the dedicated sprite rendering path. Note that, in that follow-up patch, order to see the changes in action, I had to change `bevymark` to avoid using the same seed for all its parallel RNGs. Without that change, the fact that each RNG uses the same seed means that each texture effectively gets its own Z layer, causing the benchmark to degenerate into the optimal scenario for batching in the non-bindless case. The follow-up PR scrambles the RNG seeds to avoid this pathological, unrealistic behavior.

pcwalton · 2026-07-02T04:44:15Z

check-advisories failure seems unrelated to this PR.

Zeophlite · 2026-07-02T06:19:04Z

@@ -238,7 +237,7 @@ pub struct AlphaMask2dBinKey {
    /// the ID of another type of asset.
    pub asset_id: UntypedAssetId,
    /// The ID of a bind group specific to the material.


nit: the comment should refer to an index, not ID

Zeophlite · 2026-07-02T06:20:42Z

+
+/// A [`Plugin`] that supplies generic infrastructure for 2D materials.
+#[derive(Default)]
+pub struct Materials2dPlugin;


nit: Material2dPlugin

Zeophlite

Thanks for taking the time to improve 2d

pcwalton requested review from IceSentry and kfc35 July 1, 2026 22:23

pcwalton added the A-Rendering Drawing game state to the screen label Jul 1, 2026

github-project-automation Bot added this to Rendering Jul 1, 2026

github-project-automation Bot moved this to Needs SME Triage in Rendering Jul 1, 2026

pcwalton added the C-Performance A change motivated by improving speed, memory usage or compile times label Jul 1, 2026

pcwalton force-pushed the bind-group-allocator-2d branch from 87906a6 to c889703 Compare July 1, 2026 22:25

Zeophlite reviewed Jul 1, 2026

View reviewed changes

Comment thread crates/bevy_render/src/material_bind_groups.rs Outdated

Zeophlite reviewed Jul 1, 2026

View reviewed changes

pcwalton force-pushed the bind-group-allocator-2d branch from c889703 to 8561eed Compare July 2, 2026 04:34

pcwalton added 2 commits July 1, 2026 21:35

Address review comment

23f95de

Update examples/README.md

6c88dc4

pcwalton changed the title ~~Make the built-in 2D materials bindless.~~ Add bindless support for 2D materials. Jul 2, 2026

Allowlist RenderMaterial2dInstances for ambiguity detection

f409ad2

Zeophlite reviewed Jul 2, 2026

View reviewed changes

Zeophlite approved these changes Jul 2, 2026

View reviewed changes

Zeophlite added the S-Needs-Review Needs reviewer attention (from anyone!) to move forward label Jul 2, 2026

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Uh oh!

Add bindless support for 2D materials.#24833

Add bindless support for 2D materials.#24833
pcwalton wants to merge 4 commits into
bevyengine:mainfrom
pcwalton:bind-group-allocator-2d

pcwalton commented Jul 1, 2026 •

edited

Loading

Uh oh!

Uh oh!

Zeophlite Jul 1, 2026

Uh oh!

pcwalton Jul 2, 2026

Uh oh!

Zeophlite Jul 2, 2026

Uh oh!

github-actions Bot commented Jul 1, 2026

Uh oh!

pcwalton commented Jul 2, 2026

Uh oh!

Zeophlite Jul 2, 2026

Uh oh!

Zeophlite Jul 2, 2026

Uh oh!

Zeophlite left a comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Uh oh!

Conversation

pcwalton commented Jul 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

Zeophlite Jul 1, 2026

Choose a reason for hiding this comment

Uh oh!

pcwalton Jul 2, 2026

Choose a reason for hiding this comment

Uh oh!

Zeophlite Jul 2, 2026

Choose a reason for hiding this comment

Uh oh!

github-actions Bot commented Jul 1, 2026

Uh oh!

pcwalton commented Jul 2, 2026

Uh oh!

Zeophlite Jul 2, 2026

Choose a reason for hiding this comment

Uh oh!

Zeophlite Jul 2, 2026

Choose a reason for hiding this comment

Uh oh!

Zeophlite left a comment

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

pcwalton commented Jul 1, 2026 •

edited

Loading