[Mlir-commits] [mlir] [AMDGPU] Implement gpu.subgroup_reduce with DPP intrinsics on AMD GPUs (PR #133204)

[Krzysztof Drewniak]

https://github.com/krzysz00 commented:

Overall:

We need - and this is probably calling out for making a separate PR just for permlanex16 - tests in `mlir/test/Dialect/LLVMIR/rocdl.mlir` and `mlir/test/Target/LLVMIR/rocdl.mlir` for `permlanex16`

Looking at the Nvidia code, they're testing their equivalent of this pattern using the transform dialect - see `mlir/test/Conversion/GPUToNVVM/gpu-to-nvvm.mlir` . While we aren't doing a conversion pattern and so we can't just copy of them, defining a transform dialect op for these rewrites may be a good idea.

https://github.com/llvm/llvm-project/pull/133204