wgpu: Validation Error Due to Buffer Binding Size Mismatch

[discordance]

Description

When running this Example with the latest commit of burn on master, a wgpu validation error is raised during the backward pass:

wgpu error: Validation Error

Caused by:
    In ComputePass::end
    In a dispatch command, indirect:false
    Buffer is bound with size 4 where the shader expects 8 in group[0] compact index 0

It appears that the error is caused by an updated minimum alignment requirement in wgpu according to @wingertge

Forward pass behave as expected.

You can reproduce the bug cloning This simple example, using the wgpu branch.

[laggui]

As mentioned on discord, this is actually a side effect of the issue #2776.

Let me know if you still see the issue, otherwise will close when the repeat dim backward PR is merged 🙂

[discordance]

Okay, testing here, I have a new one now:

thread 'main' panicked at ~/.cargo/registry/src/index.crates.io-6f17d22bba15001f/wgpu-24.0.1/src/backend/wgpu_core.rs:1079:26:
wgpu error: Validation Error

Caused by:
In Device::create_bind_group_layout
Too many bindings of type StorageBuffers in Stage ShaderStages(COMPUTE), limit is 31, count was 32. Check the limit max_storage_buffers_per_shader_stage passed to Adapter::request_device

[laggui]

Linking #1970

Guess there might still be an issue with fusion. If you try without the feature (enabled by default), do you still hit the limit?

[discordance]

It works when fusion is disabled
when fusion is enabled, repeat operation makes it crash.