Make wgpu compiler dynamic and determined by adapter #470
Conversation
| // Minimum supported | ||
| max_shared_memory_size: 49152, |
There was a problem hiding this comment.
it used to just be hardcoded to that for both CUDA and HIP, so I assume it was the minimum supported? I now check the actual minimum for CUDA (since it's almost always higher than 48k), but I don't fully understand the HIP code and can't really test if my code works so I used the hardcoded constant from before.
| compilation_options: &Self::CompilationOptions, | ||
| mode: ExecutionMode, | ||
| ) -> Self::Representation { | ||
| match self { |
There was a problem hiding this comment.
I guess you'll need an extra allow for clippy when spirv isn't enabled.
There was a problem hiding this comment.
not sure what you mean, I don't see any clippy warnings on either CI or locally with SPIR-V disabled
|
Since this touches relevant code parts I'll mention this issue: #355 Would be nice if possible to get hold of a client when doing device sharing. |
This PR doesn't touch the part that selects an adapter (which is where it errors), so I'd put that in a separate PR. |
Makes the
cubecl-wgpucompiler dynamic and automatically checks the backend on initialization.Breaking changes
Manually compiling kernels requires a signature change from
Compiler::compile(...)toCompiler::default().compile(...).