[Mlir-commits] [mlir] [MLIR][XeGPU] Add uniform values distribution pattern (PR #176737)

Artem Kroviakov llvmlistbot at llvm.org
Mon Jan 26 03:13:15 PST 2026


================
@@ -2036,14 +2106,16 @@ void xegpu::populateXeGPUSubgroupDistributePatterns(
                StoreMatrixDistribution,
                MemrefExtractAlignedPointerAsIndexDistribution>(
       patterns.getContext(),
-      /*pattern benefit=*/regularPatternBenefit);
+      /*pattern benefit=*/PatternHierarchy::Regular);
   // For following patterns, we need to override the regular vector distribution
   // patterns. Therefore, assign higher benefit.
   patterns
       .add<VectorShapeCastDistribution, VectorExtractStridedSliceDistribution,
            VectorInsertStridedSliceDistribution, VectorBroadcastDistribution>(
           patterns.getContext(),
-          /*pattern benefit=*/highPatternBenefit);
+          /*pattern benefit=*/PatternHierarchy::AboveRegular);
+  patterns.add<SinkUniformOps>(patterns.getContext(),
+                               /*pattern benefit=*/PatternHierarchy::High);
----------------
akroviakov wrote:

I agree that if xegpu-specific vector ops distribution patterns fail if no layouts are present, then we can place uniform ops pattern alongside, without introducing a new priority level.

https://github.com/llvm/llvm-project/pull/176737


More information about the Mlir-commits mailing list