[Mlir-commits] [mlir] [mlir][XeGPU] add WgToSg distribution pattern for load_matrix and store_matrix. (PR #154403)

[Igor Zamyatin]

================
@@ -723,8 +752,8 @@ struct WgToSgElementwiseOp : public ConversionPattern {
 // is lowered to:
 //   #a = #xegpu.layout<inst_data = [16, 16]>
 //   #b = #xegpu.layout<inst_data = [8, 16]>
-//   store_matrix %1, %slm <{layout_input_0 = #a}> : vector<32x16>, mem_desc<32x64xf32>
-//   %d = load_matrix %slm <{layout_result_0 = #a}> : mem_desc<32x64xf32> -> vector<16x32xf32>
+//   store_matrix %1, %slm <{layout_input_0 = #a}> : vector<32x16>, matrix_desc<32x64xf32>
----------------
Garra1980 wrote:

why matrix_desc here?

https://github.com/llvm/llvm-project/pull/154403