[Mlir-commits] [mlir] [mlir][gpu] Disjoint patterns for lowering clustered subgroup reduce (PR #109158)

[Andrea Faulds]

================
@@ -210,13 +210,25 @@ Value createSubgroupShuffleReduction(OpBuilder &builder, Location loc,
 struct ScalarSubgroupReduceToShuffles final
     : OpRewritePattern<gpu::SubgroupReduceOp> {
   ScalarSubgroupReduceToShuffles(MLIRContext *ctx, unsigned subgroupSize,
-                                 unsigned shuffleBitwidth,
+                                 unsigned shuffleBitwidth, bool matchClustered,
                                  PatternBenefit benefit)
       : OpRewritePattern(ctx, benefit), subgroupSize(subgroupSize),
-        shuffleBitwidth(shuffleBitwidth) {}
+        shuffleBitwidth(shuffleBitwidth), matchClustered(matchClustered) {}
 
   LogicalResult matchAndRewrite(gpu::SubgroupReduceOp op,
                                 PatternRewriter &rewriter) const override {
+    if (op.getClusterSize().has_value() != matchClustered) {
+      if (matchClustered) {
+        return rewriter.notifyMatchFailure(
+            op, "op is non-clustered but pattern is configured to only match "
+                "clustered ops");
+      } else {
+        return rewriter.notifyMatchFailure(
+            op, "op is clustered but pattern is configured to only match "
+                "non-clustered ops");
+      }
----------------
andfau-amd wrote:

I don't really like the asymmetry in this version. I've pushed another approach, does it look better?

https://github.com/llvm/llvm-project/pull/109158