[Mlir-commits] [mlir] [MLIR] Add apply_patterns.vector.arm_neon.contraction_to_i8mm TD Op (PR #140251)

Momchil Velikov llvmlistbot at llvm.org
Thu May 22 07:08:57 PDT 2025


momchil-velikov wrote:

> > In that case the TD op has to be simply apply_patterns.vector.lower_contraction (which already exists).
> 
> Why not `apply_patterns.arm_neon.lower_contraction`?

Because it's less descriptive - there's no indication on which dialect it operates. It also pollutes the "apply_patterns" namespace (considering "lower_contraction" is a generic name that does not allow unprincipled  dubious implications like "oh, we have contraction only on vectors, so it must be operating on the vector dialect").

Why `apply_patterns.arm_neon.lower_contraction`?

Why not `apply_patterns.vector.arm_neon.*`  and `apply_patterns.vector.arm_sve.*` ?


https://github.com/llvm/llvm-project/pull/140251


More information about the Mlir-commits mailing list