Commit b858309d authored by Jacques Pienaar's avatar Jacques Pienaar
Browse files

[mlir] Only attempt to vectorize conv if conv. 

Avoids hitting assertions due to unsupported convolution patterns.

See https://github.com/openxla/iree/issues/15207#issuecomment-1767650797
parent d6a92611

mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp

+10 −5
Original line number Diff line number Diff line
@@ -1693,6 +1693,7 @@ LogicalResult mlir::linalg::vectorize(RewriterBase &rewriter, Operation *op, 
          .Case<linalg::LinalgOp>([&](auto linalgOp) {

            // TODO: isaConvolutionOpInterface that can also infer from generic

            // features. Will require stride/dilation attributes inference.

            if (isa<ConvolutionOpInterface>(linalgOp.getOperation())) {

              FailureOr<Operation *> convOr =

                  vectorizeConvolution(rewriter, linalgOp);

              if (succeeded(convOr)) {
@@ -1700,6 +1701,10 @@ LogicalResult mlir::linalg::vectorize(RewriterBase &rewriter, Operation *op, 
                return success();

              }



              LDBG("Unsupported convolution can't be vectorized.\n");

              return failure();

            }



            LDBG("Vectorize generic by broadcasting to the canonical vector "

                 "shape\n");