feat(CUDA): add new target for testing

[chrsmcgrr]

We want to try a simplified GPU pipeline for CUDA and test on top of it

This adds a new cellar-cuda target backend that implements a simplified CUDA pipeline. The key changes include:

• A new CUDA CodeGen library that will be used to house our new pipeline for CUDA.
• CUDAKernelConfig contains all the current strategies for lowering down to PTX. We will have for now TileAndFuse, VectorDistribute and GPUDistribute as our main three strategies for now.

Most of the code is copied from the original target. I would say review and check how everything is structured rather than looking at the details themselves.

[chrsmcgrr]

• feat(CUDA): add new target for testing #30 👈 (View in Graphite)
• integrate-iree-20250217

This stack of pull requests is managed by Graphite. Learn more about stacking.

[ziereis]

Im still a little confused about what pipelines we want to run now.

The KernelConfig i could find the following getting set:

CodeGenPipeline::LLVMGPUMatmulTensorCore
CodeGenPipeline::LLVMGPUMatmulTensorCoreMmaSync
CodeGenPipeline::LLVMGPUVectorDistribute
CodeGenPipeline::LLVMGPUPadAndVectorDistribute
CodeGenPipeline::LLVMGPUTileAndFuse
CodeGenPipeline::LLVMGPUDistribute
CodeGenPipeline::LLVMGPUVectorize
CodeGenPipeline::LLVMGPUBaseLowering

my assumption was we only want:

CodeGenPipeline::LLVMGPUTileAndFuse
CodeGenPipeline::LLVMGPUDistribute
CodeGenPipeline::LLVMGPUVectorize

Im also pretty sure that these two pipelines dont work for nvidia (but i could be wrong)

CodeGenPipeline::LLVMGPUVectorDistribute
CodeGenPipeline::LLVMGPUPadAndVectorDistribute

[chrsmcgrr]

@ziereis so we decided to go for LLVMGPUTileAndFuse, LLVMGPUVectorDistribute, LLVMGPUDistribute as per our ticket. If they don't work the idea is we figure out why and fix them :D

[ziereis]

i thought we agreed on not using LLVMGPUVectorDistribute as its super complex and more or less "deprecated" as upstream IREE also wants to move to only using TileAndFuse for matmuls (e.g iree-org#19854, https://github.com/iree-org/iree/blob/00e88733e6b8c8cdb351d4516509f56daebdf604/compiler/src/iree/compiler/Codegen/LLVMGPU/KernelConfig.cpp#L2391)

i might be misunderstanding something but it thought the main goal was to use TileAndFuse for almost everything

Also in this case now we still have 4 pipeline that are supposed to handle matmuls:

CodeGenPipeline::LLVMGPUMatmulTensorCore
CodeGenPipeline::LLVMGPUMatmulTensorCoreMmaSync
CodeGenPipeline::LLVMGPUTileAndFuse
CodeGenPipeline::LLVMGPUVectorDistribute

[chrsmcgrr]

I don't mind reducing the complexity by sticking to 2 pipelines (we need GPUDistribute as a fallback). These pipelines are starting points we will probably hard-fork them as well and create our own.

[maxbartel]

Nice first step! Looking even more into it, forking is the only sane choice. We do quite a radical trimming