[all-commits] [llvm/llvm-project] b414d7: [MLIR][NVVM] Unify and move to a single tcgen05_mm...

[Kirill Vedernikov via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: b414d7795d2965ffaa9457a7e56dca76fd07cce1
      https://github.com/llvm/llvm-project/commit/b414d7795d2965ffaa9457a7e56dca76fd07cce1
  Author: Kirill Vedernikov <kvedernikov at nvidia.com>
  Date:   2026-03-05 (Thu, 05 Mar 2026)

  Changed paths:
    M mlir/include/mlir/Dialect/LLVMIR/NVVMOps.td
    M mlir/lib/Dialect/LLVMIR/IR/NVVMDialect.cpp
    M mlir/test/Target/LLVMIR/nvvm/tcgen05-mma-block-scale-shared.mlir
    M mlir/test/Target/LLVMIR/nvvm/tcgen05-mma-block-scale-tensor.mlir
    M mlir/test/Target/LLVMIR/nvvm/tcgen05-mma-invalid.mlir
    M mlir/test/Target/LLVMIR/nvvm/tcgen05-mma-sp-block-scale-shared.mlir
    M mlir/test/Target/LLVMIR/nvvm/tcgen05-mma-sp-block-scale-tensor.mlir

  Log Message:
  -----------
  [MLIR][NVVM] Unify and move to a single tcgen05_mma_kind attr for all tcgen05.mma Ops (#184433)

This change unifies using of `tcgen05_mma_kind` attribute for
tcgen05.mma Ops in MLIR.

Before this change there were two block scale attributes used for
tcgen05.mma Ops. One was `MMABlockScaleKindAttr` with `mxf8f6f4`, `mxf4`
and `fxf4nvf4` values used for `tcgen05.mma.block_scale` and
`tcgen05.mma.sp.block_scale`. Another one was `Tcgen05MMAKindAttr` with
`f16`, `tf32`, `f8f6f4` and `i8` values used for `tcgen05.mma`,
`tcgen05.mma.sp`, `tcgen05.mma.ws` and `tcgen05.mma.ws.sp`.

`Tcgen05MMAKindAttr` has been extended with values from
`MMABlockScaleKindAttr`. Now there is `tcgen05_mma_kind` attribute only
for all `tcgen05.mma` Ops in MLIR.

Backward compatibility is not supported. Existing tests and scripts
should be updated to use `tcgen05_mma_kind` attribute instead of
`block_scale_kind` for all tcgen05.mma MLIR Ops.



To unsubscribe from these emails, change your notification settings at https://github.com/llvm/llvm-project/settings/notifications