[metal] Metal compute shader passthrough

[syl20bnr]

Description
This PR adds the possibility to directly pass a metal shader and create a compute pipeline to execute the passed shader without any validation. This allows CubeCL to use its wgpu runtime to execute metal kernels directly compiled by CubeCL JIT compiler and benefit from support for advanced metal features.

Testing
Tested with a WIP metal compiler using the wgpu runtime of CubeCL.
More info about how to test this soon.

Squash or Rebase?

Ok to be squashed.

Checklist

• Run cargo fmt.
• Run taplo format.
• Run cargo clippy. If applicable, add:
  • --target wasm32-unknown-unknown
• Run cargo xtask test to run tests.
• If this contains user-facing changes, add a CHANGELOG.md entry.

[cwfitzgerald]

It would be great to get some description of the shader ABI we present to shaders and how things get translated. It doesn't need to be as strict as in wgpu-hal, but should guide people on what wgpu expects the shader to have.

[DJMcNab]

(Linking to #3103)

[syl20bnr]

@cwfitzgerald Thank you for the review.

I think we should unify this entrypoint with create_shader_module_spirv and call it create_shader_module_passthrough which takes a ShaderModulePassthrough enum with each descriptor as needed.

I pushed a simple change for this, is that what you had in mind ?

It would be great to get some description of the shader ABI we present to shaders and how things get translated. It doesn't need to be as strict as in wgpu-hal, but should guide people on what wgpu expects the shader to have.

Not sure what do you mean, do you have an example of such description ?

Also can you point me to where I could add a test for the metal passthrough if needed ?

[cwfitzgerald]

Not sure what do you mean, do you have an example of such description ?

Take a look at https://wiki.libsdl.org/SDL3/SDL_CreateGPUShader#remarks which talks about the order that SDL requires the shader bindings to be, depending on the backend. Ideally we would have something like this for MSL as its not always straight forward, but we can punt this to another PR if it would be too much hassle for now.

Also can you point me to where I could add a test for the metal passthrough if needed ?

Make sure you update the PR to latest, but then tests/tests/wgpu-gpu has gpu-enabled tests. You can write a test which is skipped on all platforms other that metal, such as

wgpu/tests/tests/wgpu-gpu/oob_indexing.rs

Line 242 in 2c777f7

.skip(FailureCase::backend(Backends::all() - Backends::DX12)),

defines a test which operates on all platforms except DX12.

[cwfitzgerald]

Looking good, some comments

[cwfitzgerald]

WebAssembly is also failing due to missing methods in wgpu/src/backend/webgpu

[cwfitzgerald]

Alright, I'll drive this over the line, thanks!

[LegNeato]

@syl20bnr @cwfitzgerald Pretty sure this broke spirv passthrough: #7502

[syl20bnr]

@LegNeato do you have more info ?

I tested with our SpirV compiler which used the pass-through and it seems to work fine.

Oh OK just saw your issue. Looking into it. We indeed don't use this macro.

Edit: thanks for the fix 😁