[Mlir-commits] [mlir] [MLIR][XeGPU] Add handling for unit-dim expansion in ShapeCast workgroup-to-subgroup distribution (PR #171758)

[Jianhui Li]

================
@@ -1530,12 +1530,12 @@ struct VectorBroadcastDistribution : public gpu::WarpDistributionPattern {
       if (rankDiff == 0) {
         SetVector<int64_t> broadcastUnitDims =
             broadcastOp.computeBroadcastedUnitDims();
-        resultLayout = resultLayout.setUnitDimData(broadcastUnitDims);
         bool isEqualTo = sourceLayout.isEqualTo(resultLayout);
         if (!isEqualTo)
           return rewriter.notifyMatchFailure(
               warpOp, "For same-rank broadcast, source must be identical to "
                       "adjusted result layouts with unit dims.");
+        resultLayout = resultLayout.setUnitDimData(broadcastUnitDims);
----------------
Jianhui-Li wrote:

This one is set sg_data/lane_data/inst_data, and the other setUnitDimLayout() is set sg_layou and lane_layout

https://github.com/llvm/llvm-project/pull/171758