[Mlir-commits] [mlir] [MLIR][Linalg] Scalable Vectorization of Reduction (PR #97788)

[Zhaoshi Zheng]

================
@@ -1778,6 +1784,9 @@ vectorizeDynamicLinalgOpPrecondition(linalg::LinalgOp op,
   if (isa<ConvolutionOpInterface>(op.getOperation()))
     return vectorizeDynamicConvOpPrecondition(op, flatten1DDepthwiseConv);
 
+  if (isLinalgReduction(op))
+    return reductionPreconditions(op);
+
----------------
zhaoshiz wrote:

linalg.reduce ops will fail the next check L1792~L1795 and cause vectorizeLinalgOpPrecondition() to return failure.
For static-shaped reduce ops I don't think we need it. But after tiling with scalable vector like [[4]], we got dynamic-shaped ops.


https://github.com/llvm/llvm-project/pull/97788