[Mlir-commits] [mlir] [MLIR][XeGPU][VectorToXeGPU] Fixed transposed transfer_read/write (PR #197281)

[Jianhui Li]

================
@@ -288,7 +292,7 @@ computeMemrefMeta(OpType xferOp, PatternRewriter &rewriter) {
 //   %offsets =  memref_offset + orig_offset + local_offsets
 static Value computeOffsets(VectorTransferOpInterface xferOp,
----------------
Jianhui-Li wrote:

Instead of change two function signature and pass two strides to computeOffsets(), you could do the following: 
  1. computeMemrefMeta() no longer permutes
  2. Inside computeOffsets(), it builds a local permutedStrides copy (via
  adjustStridesForPermutation) used only for the local-offset step-vector loop; the base-offset loop keeps the original memref-order strides.

https://github.com/llvm/llvm-project/pull/197281