llvm-project

76699fb9 - [MLIR][NVVM][NVPTX] Support for new mma/mma.sp variants from PTX 9.1 (#182325)

[MLIR][NVVM][NVPTX] Support for new mma/mma.sp variants from PTX 9.1 (#182325) This change adds support for `.scale_vec::4X` with `.ue8m0` as `.stype` with `.kind::mxf4nvf4` for `mma/mma.sp` instructions introduced in [PTX ISA 9.1](https://docs.nvidia.com/cuda/parallel-thread-execution/index.html?highlight=mma%2520sp#ptx-isa-version-9-1). Also, it updates MLIR mma/mma.sp block scale tests with struct usage instead of vector.