Upgrade to wgpu 24 #436
Upgrade to wgpu 24 #436
Conversation
ArthurBrussee
commented
Jan 19, 2025
ArthurBrussee
commented
- Add atomic floats feature matrix to wgpu. In the future when f16 etc. are supported on wgpu this feature matrix might need to be more granular.
- Change checked execution - infinite loops are now again checked in checked mode as they can cause UB. It might make sense to make more granular options for this later.
crates/cubecl-wgpu/src/runtime.rs
Outdated
| if features.contains(wgpu::Features::SHADER_FLOAT32_ATOMIC) { | ||
| device_props.register_feature(Feature::AtomicFloat(AtomicFeature::LoadStore)); | ||
| device_props.register_feature(Feature::AtomicFloat(AtomicFeature::Add)); | ||
| device_props.register_feature(Feature::AtomicFloat(AtomicFeature::MinMax)); |
There was a problem hiding this comment.
The way I've implemented the features for float atomics, AtomicFeature should be combined with registering the relevant atomic types. As far as I can tell, if an atomic type is supported at all then all operations are supported on it.
So this should also register Feature::Type(Elem::AtomicFloat(FloatKind::F32))
There was a problem hiding this comment.
Also, as far as I can tell MinMax isn't supported for float at least not universally, it isn't available in Vulkan on my card either, and CUDA doesn't have support for it at all.
There was a problem hiding this comment.
The PR feature matrix for float atomics says it's supported in Vulkan (but not on the 3080), but not implemented at all in wgpu. So it definitely should be disabled.
There was a problem hiding this comment.
Ah bad memory really thought the PR said it was supported, nvm.
And I see about the Elem::AtomicFloat feature, that makes sense! Thank you
| ) -> Arc<ComputePipeline> { | ||
| let source = &kernel.source; | ||
| // Cube always in principle uses unchecked modules. Certain operations like | ||
| // indexing are instead checked by cube. The WebGPU specification only makes | ||
| // incredibly loose guarantees that Cube can't rely on. Additionally, kernels | ||
| // can opt in/out per operation whether checks should be performed which can be faster. | ||
| // | ||
| let checks = ShaderRuntimeChecks { | ||
| bounds_checks: false, |
There was a problem hiding this comment.
I would add a comment that automatic bounds checks are done by the cubecl compiler.
There was a problem hiding this comment.
The comment above said as much but clarified this a bit
