MacOS Rendering Corruption starting on 0.19.0

[snappa-app]

Bevy version and features

• Bevy 0.19.1, 0.19.0 (not reproducible in 0.18.1)
• Default features

[Optional] Relevant system information

If your bug is rendering-related, copy the adapter info that appears when you run Bevy.

AdapterInfo { name: "Apple M3 Pro", vendor: 0, device: 0, device_type: IntegratedGpu, device_pci_bus_id: "", driver: "", driver_info: "", backend: Metal, subgroup_min_size: 4, subgroup_max_size: 64, transient_saves_memory: true }

What you did

I'm encountering a combination of 2 issues:

• Materials that swapping around to different meshes
• Meshes flickering in and out of visibility, every other frame

It happens under these circumstances, but I haven't been able to narrow it down specifically:

• Only impacts 0.19.x versions
• Impacts macOS, does not impact Windows
• Only happens with GPU Preprocessing - I think this is a strong clue about the root cause, see the bottom for more information
• The problem gets worse using:
  • GLTF Meshes (I'm relying on instancing a lot)
  • FXAA instead of MSAA

I'm working really hard to get a minimal reproducible build, But I haven't been successful. The problem seems to scale with complexity, so I think you need a complicated rendering setup to get the issue to trigger. I would be willing to share my reproduction with select maintainers if requested.

What went wrong

Here's a video of the corruption:

recordning.mp4

Additional information

Theory About GPU Preprocessing

I'm able to fix the issue by forcing the WGPU to use CPU preprocessing. This changes the log GPU preprocessing is fully supported on this device. to (GPU preprocessing is not supported on this device. Falling back to CPU preprocessing.),. When I do this, my framerate improves, and the flickering and material swapping goes away entirely.

Hunches: I am not super knowledgeable about WGPU internals, but doing some research, here's some thoughts:

• I tried setting WgpuSettings::disabled_features(WgpuFeatures::TEXTURE_BINDING_ARRAY), thinking the issue was related to bindless textures, but this didn't solve it
• I think the issue lies in the GPU indirect drawing path that GPU preprocessing/culling turns on.
• I'm thinking this could be related to Indirect drawing causes AlphaMode::Blend meshes to flicker #21739, AlphaMode::Blend meshes flickering with indirect drawing
• [I did some AI-assisted digging here] There was a change between 0.18 and 0.19 where MeshAllocator's slab-resize still does a queue.submit that never joined the new batched PendingCommandBuffers model everything else in the render graph now goes through, so there's one remaining out-of-band submission in an otherwise fully-batched frame

Here's the workaround that fixes the issue even on 0.19.1:

  let render_plugin = bevy::render::RenderPlugin {
    render_creation: macos_wgpu_settings_forcing_cpu_preprocessing().into(),
    ..default()
  };
app.add_plugins((DefaultPlugins.set(render_plugin),...));

...

// Hack to disable GPU preprocessing on macOS
#[cfg(target_os = "macos")]
fn macos_wgpu_settings_forcing_cpu_preprocessing() -> bevy::render::settings::WgpuSettings {
  use bevy::render::settings::WgpuLimits;

  bevy::render::settings::WgpuSettings {
    constrained_limits: Some(
      WgpuLimits {
        max_binding_array_elements_per_shader_stage: u32::MAX,
        max_binding_array_sampler_elements_per_shader_stage: u32::MAX,
        ..WgpuLimits::downlevel_webgl2_defaults()
      }
      .using_resolution(WgpuLimits::default()),
    ),
    ..default()
  }
}