[Mlir-commits] [mlir] [mlir][vector] Add pattern to break down small reductions into arith ops (PR #75727)

[Andrzej Warzyński]

================
@@ -1578,6 +1580,50 @@ struct ReduceRedundantZero final : OpRewritePattern<vector::ReductionOp> {
   }
 };
 
+struct BreakDownVectorReduction final : OpRewritePattern<vector::ReductionOp> {
+  BreakDownVectorReduction(MLIRContext *context,
+                           unsigned maxNumElementsToExtract,
+                           PatternBenefit benefit)
+      : OpRewritePattern(context, benefit),
+        maxNumElementsToExtract(maxNumElementsToExtract) {}
+
+  LogicalResult matchAndRewrite(vector::ReductionOp op,
+                                PatternRewriter &rewriter) const override {
+    VectorType type = op.getSourceVectorType();
+    if (type.isScalable() || op.isMasked())
+      return failure();
----------------
banach-space wrote:

Could you add a test for scalable vectors and masked ops? I will often rely on tests to gauge the level of support.

https://github.com/llvm/llvm-project/pull/75727