[Mlir-commits] [mlir] [mlir][vector] Remove MatrixMultiplyOp and FlatTransposeOp from Vector dialect (PR #144307)

[Kunwar Grover]

Groverkss wrote:

LGTM on removing the ops, but I don't know if we can introduce a llvm dialect dependency like this. Someone going down the Vector -> SPIRV path might get suprised by this. I don't like churn too, but i think we need to move this down to conversions.

Maybe we can split this into two patches, one that changes to conversion directly to llvm intrinsic and removes the old transformation and then the op removal?

I'm ok with any way, as long as we also ensure we don't suprise anyone trying to use a backend other than LLVM.

https://github.com/llvm/llvm-project/pull/144307