[Mlir-commits] [mlir] [mlir][linalg] Implement Conv2D using Winograd Conv2D algorithm (PR #96181)

[llvmlistbot at llvm.org]

================
@@ -2734,6 +2734,113 @@ FailureOr<SmallVector<Value>> SoftmaxOp::decomposeOperation(OpBuilder &b) {
   return SmallVector<Value>{result};
 }
 
+//===----------------------------------------------------------------------===//
----------------
Max191 wrote:

These verifiers will not work for dynamic shapes. Can you support dynamic cases? The transform is only supported for static shapes right now, but shapes can become dynamic when tiling.

You can create an expected output shape from the input, allowing dynamic dims, and compare with the actual output shape. This helper may be useful: https://github.com/llvm/llvm-project/blob/586114510c5fa71d1377c7f53e68a3b12c472aa2/mlir/lib/IR/TypeUtilities.cpp#L58

This way will also make it easy to check that the batch/channel dimensions match for the input and output.

https://github.com/llvm/llvm-project/pull/96181