Support `transpose` mode for `gpu.subgroup` WMMA ops
authorNavdeep Katel <navdeep@polymagelabs.com>
Mon, 5 Dec 2022 10:44:56 +0000 (16:14 +0530)
committerUday Bondhugula <uday@polymagelabs.com>
Mon, 5 Dec 2022 17:07:02 +0000 (22:37 +0530)
commit3d35546cd1680b0e087fb6c9976799760146c377
tree11bdf50bc00394e7e9d900ae0bbde196b0913d33
parent03b3017aef14efa705d67767d229752b1c6d0b2d
Support `transpose` mode for `gpu.subgroup` WMMA ops

Add support for loading, computing, and storing `gpu.subgroup` WMMA ops
in transpose mode as well. Update the GPU to NVVM lowerings to support
`transpose` mode and update integration tests as well.

Reviewed By: ThomasRaoux

Differential Revision: https://reviews.llvm.org/D139021
mlir/include/mlir/Dialect/GPU/IR/GPUOps.td
mlir/lib/Conversion/GPUToNVVM/WmmaOpsToNvvm.cpp
mlir/lib/Conversion/GPUToSPIRV/WmmaOpsToSPIRV.cpp
mlir/lib/Conversion/VectorToGPU/VectorToGPU.cpp
mlir/test/Conversion/GPUToNVVM/wmma-ops-to-nvvm.mlir
mlir/test/Conversion/GPUToSPIRV/wmma-ops-to-spirv.mlir
mlir/test/Integration/GPU/CUDA/TensorCore/wmma-matmul-f16.mlir