[Mlir-commits] [mlir] [MLIR][XeGPU][XeVM] Add lowering for load_nd with MX scale element type. (PR #198020)

[Artem Kroviakov]

https://github.com/akroviakov approved this pull request.

Functionally looks fine and is actually a good base for a "generic" load_nd/store_nd ops implementation for hw that does not support 2d block loads.
Using lane_id for further offset manipulation in conversion looks hacky, but so is the whole pattern, so it should be fine for the initial enabling.

https://github.com/llvm/llvm-project/pull/198020