[Mlir-commits] [mlir] [MLIR][XeGPU] Add distribution pattern for xegpu.load & store for sg to wi pass (PR #181917)

[Jianhui Li]

================
@@ -395,6 +395,71 @@ struct SgToWiPrefetchNd : public OpConversionPattern<xegpu::PrefetchNdOp> {
   }
 };
 
+/// Distributes a subgroup-level LoadGather (xegpu.load) op to workitem-level.
+struct SgToWiLoadGather : public OpConversionPattern<xegpu::LoadGatherOp> {
+  using OpConversionPattern<xegpu::LoadGatherOp>::OpConversionPattern;
+
+  LogicalResult
+  matchAndRewrite(xegpu::LoadGatherOp op, OpAdaptor adaptor,
+                  ConversionPatternRewriter &rewriter) const override {
+    xegpu::DistributeLayoutAttr layout = op.getAnchorLayout();
+    if (!layout)
+      return failure();
+
+    VectorType resultTy = op.getValueType();
+    if (!resultTy)
+      return failure();
+
+    // Check that leading dimensions are unit.
+    int chunkSize = op.getChunkSize().value_or(1);
+    int effectiveVecRank = (chunkSize == 1) ? 1 : 2;
+    ArrayRef<int64_t> shape = resultTy.getShape();
+    if (llvm::any_of(shape.take_front(resultTy.getRank() - effectiveVecRank),
+                     [](int64_t d) { return d != 1; }))
+      return rewriter.notifyMatchFailure(
+          op, "Only unit dimensions allowed for the leading "
+              "dimensions of the load vector!");
+
+    auto expectedWiResultTyOrFailure =
+        xegpu::getDistVecTypeBasedOnLaneLayout(layout, resultTy);
+    if (failed(expectedWiResultTyOrFailure))
+      return rewriter.notifyMatchFailure(
+          op,
+          "unable to compute expected workitem vector type from lane layout");
+
+    VectorType expectedWiResultTy = expectedWiResultTyOrFailure.value();
+    VectorType supportedWiResultTy =
+        VectorType::get({expectedWiResultTy.getNumElements()},
+                        expectedWiResultTy.getElementType());
+
+    // Flatten offsets and mask to 1D to match the 1D result type.
+    Value offsets = adaptor.getOffsets();
----------------
Jianhui-Li wrote:

Can we use "dist_" prefix for these value get from adaptor, and "orig_" prefix for those values from op? 

The current coding require reviewer to look back to find out where it is from, so harder to maintain/read. 

https://github.com/llvm/llvm-project/pull/181917