[Mlir-commits] [mlir] [MLIR][XeGPU] Enable WG-level mxfp GEMM via generalized shape_cast collapse inference (PR #201496)

[Sang Ik Lee]

================

----------------
silee2 wrote:

The function does not do a full GEMM.
In the code, only M and N dimensions are split and distributed at WG level.
So, K dimension should be the full length for WG kernel. 
Otherwise, the sole dpas_mx op below does not cover the full K dimension.
See https://github.com/llvm/llvm-project/pull/201680 as the test there covers the full K dimension.
M and N are smaller and a K loop is added to the code but can be used as a reference point.

https://github.com/llvm/llvm-project/pull/201496