GEMM lowering paths

[jungpark-mlir]

Hi all,

I'm wondering if you plan to implement passes converting from/to gpu.subgroup_mma_compute_matrix
e.g., rock.blockwise_gemm_accel -> gpu.subgroup_mma_compute_matrix -> amdgpu.wmma

And other gpu.subgroup_mma_* operation might be also interesting to you. Especially, upstream has transforms to fold vector.contraction to gpu.subgroup_mma_compute_matrix. And gpu.subgoup_mma_compute_matrix can be converted to nvvm.wmma or spirv.NV.CooperativeMatrixMulAdd. (KHR extension will be landing soon to the spirv)

Cheers,
Jungwook

[krzysz00]

@sjw36 is working on getting a lowering from the GPU MMA ops to MFMA going.

WMMA will very much be future work there, I gather, but once the MFMA handling is landed, WMMA shouldn't be too hard (especially since you'll have our code to work off of for the register layout).

If you'd be open to reviewing https://reviews.llvm.org/D152451 , that'd be good