[mlir][gpu] Relax restriction on mma load/store op
authorThomas Raoux <thomasraoux@google.com>
Fri, 25 Mar 2022 03:31:06 +0000 (03:31 +0000)
committerThomas Raoux <thomasraoux@google.com>
Fri, 25 Mar 2022 04:03:40 +0000 (04:03 +0000)
commitd77f4836401e492fa4aee549d6f677b6ab336a81
tree9d056ab64d652d8c0f4222b7d7a6b6b9f9f74f1f
parentb62ea9b38b627ce2a96f66242c15250d59a9b134
[mlir][gpu] Relax restriction on mma load/store op

Those ops can support more complex layout as long as the most inner
dimension is contiguous.

Differential Revision: https://reviews.llvm.org/D122452
mlir/lib/Conversion/VectorToGPU/VectorToGPU.cpp
mlir/lib/Dialect/GPU/IR/GPUDialect.cpp
mlir/test/Conversion/VectorToGPU/vector-to-mma-ops.mlir
mlir/test/Dialect/GPU/invalid.mlir
mlir/test/Dialect/GPU/ops.mlir