[Mlir-commits] [mlir] [mlir][xegpu] Add initial skeleton implementation for lowering ConvertLayoutOp (PR #146176)

[Chao Chen]

================
@@ -22,7 +22,7 @@ def XeGPU_DpasResType: FixedVectorOfRankAndType<[1, 2], [XeGPU_ScalarType]>;
 def XeGPU_OffsetType: FixedVectorOfNonZeroRankOf<[Index]>;
 def XeGPU_MaskType: FixedVectorOfNonZeroRankOf<[I1]>;
 def XeGPU_ValueType: FixedVectorOfNonZeroRankOf<[XeGPU_ScalarType]>;
-def XeGPU_Vector2DType: FixedVectorOfRankAndType<[2], [XeGPU_ScalarType]>;
+def XeGPU_VectorType: VectorOfRankAndType<[1,2,3,4,5,6], [XeGPU_ScalarType]>;
----------------
chencha3 wrote:

It is inspired by a case from habana using load_gather, which used up to 5D. I was debating with myself whether to remove the rank restriction completely. 

https://github.com/llvm/llvm-project/pull/146176