[Mlir-commits] [mlir] [MLIR][XeGPU] Support vector.contract transpose_a/transpose_b via 'vector-to-gpu' patterns (PR #182885)

[Adam Siemieniuk]

adam-smnk wrote:

If it works that's fine, let's proceed as is.
But I think it introduces awkward coupling with patterns targeting GPU dialect MMA ops while we side step them and go through plain vector.

Clearly these patterns (or parts of them?) are useful in more general cases.
Consider some refactoring to expose needed vector patterns without full dependency on the MMA lowering path.
Anyway, not a blocker or anything that has to be addressed in this PR.

https://github.com/llvm/llvm-project/pull/182885