[Mlir-commits] [mlir] [MLIR][XeGPU] Improve workgroup to subgroup distribution pattern for mulit-reduction op (PR #182178)

[Artem Kroviakov]

================
@@ -1484,78 +1440,54 @@ struct WgToSgMultiDimReductionOp
       return failure();
     SmallVector<Value> sgIds = *sgIdsResult;
 
-    // Row offset: linearize reduction dimension indices
-    Value rowOffsetStore = linearizeSubgroupIndices(
-        rewriter, loc, sgIds, crossSgReductionDims, sgLayout);
-
-    // Column offset: linearize non-reduction dimension indices
-    SmallVector<int64_t> nonReductionDims;
-    for (size_t i = 0; i < sgLayout.size(); ++i) {
-      if (!llvm::is_contained(reductionDims, static_cast<int64_t>(i))) {
-        nonReductionDims.push_back(static_cast<int64_t>(i));
-      }
+    SmallVector<OpFoldResult> slmStoreOffsets;
+    for (int i = 0; i < srcVecRank; ++i) {
+      Value dimVal = sgIds[i];
+      int64_t stride =
----------------
akroviakov wrote:

Perhaps something like
```suggestion
      int64_t sgDataStride =
```
Makes it easier to follow why we do stride math for non-linearized offsets.

https://github.com/llvm/llvm-project/pull/182178