ROCm/rocMLIR

[MLIR] How to target MFMA ops for AMDGPU backend?

nirvedhmeshram opened this issue · 3 comments

nirvedhmeshram commented

I see these ops in the repository
https://github.com/ROCmSoftwarePlatform/llvm-project-mlir/blob/0ed808679fbaf19b0379f76ea2e1d3c95ac3fe88/external/llvm-project/mlir/include/mlir/Dialect/GPU/GPUOps.td#L1265
However, I dont see them available upstream. Is there currently some way to target MFMA ops for AMDGPU backend in upstream MLIR?

whchung commented

@nirvedhmeshram , we are in the process of removing AMDGPU-specific ops from GPU dialect, and move them to amdgpu / rocdl dialect. We'll upstream all the changes.

  • 👍1
krzysz00 commented

Yeah, once I get #628 merged here, I'll send it upstream

  • 👍1
nirvedhmeshram commented

Thank you!