[mlir][gpu] Disjoint patterns for lowering clustered subgroup reduce

Making the existing populateGpuLowerSubgroupReduceToShufflePatterns()
function also cover the new "clustered" subgroup reductions is proving
to be inconvenient, because certain backends may have more specific
lowerings that only cover the non-clustered type, and this creates pass
ordering constraints. This commit removes coverage of clustered
reductions from this function in favour of a new separate function,
which makes controlling the lowering much more straightforward.

Author: andfau-amd

mlir/include/mlir/Dialect/GPU/Transforms/Passes.h

@@ -73,10 +73,20 @@ void populateGpuBreakDownSubgroupReducePatterns(
 /// Collect a set of patterns to lower `gpu.subgroup_reduce` into `gpu.shuffle`
 /// ops over `shuffleBitwidth` scalar types. Assumes that the subgroup has
 /// `subgroupSize` lanes. Uses the butterfly shuffle algorithm.
+///
+/// The patterns populated by this function will ignore ops with the
+/// `cluster_size` attribute.
+/// `populateGpuLowerClusteredSubgroupReduceToShufflePatterns` is the opposite.
 void populateGpuLowerSubgroupReduceToShufflePatterns(
     RewritePatternSet &patterns, unsigned subgroupSize,
     unsigned shuffleBitwidth = 32, PatternBenefit benefit = 1);
 
+/// Disjoint counterpart of `populateGpuLowerSubgroupReduceToShufflePatterns`
+/// that only matches `gpu.subgroup_reduce` ops with a `cluster_size`.
+void populateGpuLowerClusteredSubgroupReduceToShufflePatterns(
+    RewritePatternSet &patterns, unsigned subgroupSize,
+    unsigned shuffleBitwidth = 32, PatternBenefit benefit = 1);
+
 /// Collect all patterns to rewrite ops within the GPU dialect.
 inline void populateGpuRewritePatterns(RewritePatternSet &patterns) {
   populateGpuAllReducePatterns(patterns);

mlir/lib/Dialect/GPU/Transforms/SubgroupReduceLowering.cpp

@@ -210,13 +210,24 @@ 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");
+    }
+
     auto ci = getAndValidateClusterInfo(op, subgroupSize);
     if (failed(ci))
       return failure();
@@ -262,19 +273,31 @@ struct ScalarSubgroupReduceToShuffles final
 private:
   unsigned subgroupSize = 0;
   unsigned shuffleBitwidth = 0;
+  bool matchClustered;
 };
 
 /// Lowers vector gpu subgroup reductions to a series of shuffles.
 struct VectorSubgroupReduceToShuffles final
     : OpRewritePattern<gpu::SubgroupReduceOp> {
   VectorSubgroupReduceToShuffles(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");
+    }
+
     auto ci = getAndValidateClusterInfo(op, subgroupSize);
     if (failed(ci))
       return failure();
@@ -343,6 +366,7 @@ struct VectorSubgroupReduceToShuffles final
 private:
   unsigned subgroupSize = 0;
   unsigned shuffleBitwidth = 0;
+  bool matchClustered;
 };
 } // namespace
 
@@ -358,5 +382,14 @@ void mlir::populateGpuLowerSubgroupReduceToShufflePatterns(
     RewritePatternSet &patterns, unsigned subgroupSize,
     unsigned shuffleBitwidth, PatternBenefit benefit) {
   patterns.add<ScalarSubgroupReduceToShuffles, VectorSubgroupReduceToShuffles>(
-      patterns.getContext(), subgroupSize, shuffleBitwidth, benefit);
+      patterns.getContext(), subgroupSize, shuffleBitwidth,
+      /*matchClustered=*/false, benefit);
+}
+
+void mlir::populateGpuLowerClusteredSubgroupReduceToShufflePatterns(
+    RewritePatternSet &patterns, unsigned subgroupSize,
+    unsigned shuffleBitwidth, PatternBenefit benefit) {
+  patterns.add<ScalarSubgroupReduceToShuffles, VectorSubgroupReduceToShuffles>(
+      patterns.getContext(), subgroupSize, shuffleBitwidth,
+      /*matchClustered=*/true, benefit);
 }

mlir/test/lib/Dialect/GPU/TestGpuRewrite.cpp

@@ -78,9 +78,12 @@ struct TestGpuSubgroupReduceLoweringPass
     populateGpuBreakDownSubgroupReducePatterns(patterns,
                                                /*maxShuffleBitwidth=*/32,
                                                PatternBenefit(2));
-    if (expandToShuffles)
+    if (expandToShuffles) {
       populateGpuLowerSubgroupReduceToShufflePatterns(
           patterns, /*subgroupSize=*/32, /*shuffleBitwidth=*/32);
+      populateGpuLowerClusteredSubgroupReduceToShufflePatterns(
+          patterns, /*subgroupSize=*/32, /*shuffleBitwidth=*/32);
+    }
 
     (void)applyPatternsAndFoldGreedily(getOperation(), std::move(patterns));
   }