[mlir][gpu] Add support for unsigned integer extend in vector to gpu.subgroup_mma...
authorQuinn Dawkins <quinn@nod-labs.com>
Mon, 13 Feb 2023 01:33:10 +0000 (20:33 -0500)
committerQuinn Dawkins <quinn@nod-labs.com>
Tue, 14 Feb 2023 18:09:46 +0000 (13:09 -0500)
commit5205c7126b2fea73b283e54c4e1bf409089a8d52
treec37cf6f26712c39e89a97c50ea2123665b6c0119
parentf6cd46e07fcca123620afff96675171f8fe96124
[mlir][gpu] Add support for unsigned integer extend in vector to gpu.subgroup_mma lowering

Unsigned integer types are supported in subgroup mma ops by matching
against arith.extui ops. This allows for subgroup_mma_compute ops with
mixed signedness which requires later conversions to handle this. SPIR-V
cooperative matrix ops support this while the lowering to WMMA does not.

Differential Revision: https://reviews.llvm.org/D143922
mlir/lib/Conversion/GPUToNVVM/WmmaOpsToNvvm.cpp
mlir/lib/Conversion/VectorToGPU/VectorToGPU.cpp
mlir/lib/Dialect/SPIRV/IR/SPIRVOps.cpp
mlir/test/Conversion/VectorToGPU/vector-to-mma-ops.mlir
mlir/test/Dialect/SPIRV/IR/cooperative-matrix-ops.mlir