[Mlir-commits] [mlir] [mlir][vector] Separate multi_reduce lowering into transformations, flattening, and unrolling. (PR #178974)

[Jakub Kuderski]

================
@@ -413,6 +413,60 @@ struct TwoDimMultiReductionToReduction
   }
 };
 
+/// Converts 1-D vector.multi_reduction directly to vector.reduction.
+/// This is the terminal case for unrolling - once we reach rank 1,
+/// we convert to vector.reduction which backends can optimize.
+///
+/// Example:
+/// ```mlir
+/// // Before
+/// %r = vector.multi_reduction <add>, %v, %acc [0] : vector<Nxf32> to f32
+///
+/// // After
+/// %r = vector.reduction <add>, %v, %acc : vector<Nxf32> into f32
+/// ```
+struct OneDimMultiReductionToReduction
+    : public OpRewritePattern<vector::MultiDimReductionOp> {
+  using Base::Base;
+
+  LogicalResult matchAndRewrite(vector::MultiDimReductionOp multiReductionOp,
+                                PatternRewriter &rewriter) const override {
+    auto srcRank = multiReductionOp.getSourceVectorType().getRank();
+    if (srcRank != 1)
+      return rewriter.notifyMatchFailure(multiReductionOp,
+                                         "expected source rank == 1.");
+
+    if (!multiReductionOp.isReducedDim(0))
+      return rewriter.notifyMatchFailure(
+          multiReductionOp,
+          "expected dimension 0 to be a reduction dimension.");
----------------
kuhar wrote:

also here

https://github.com/llvm/llvm-project/pull/178974