[Mlir-commits] [mlir] [mlir][linalg] Use inferConvolutionDims for generic convolution downscaling (PR #180586)

[Abhishek Varma]

================
@@ -1640,54 +1640,23 @@ FailureOr<linalg::GenericOp> deduplicateOperandsAndRemoveDeadResults(
 // functional-stye API call.
 //===----------------------------------------------------------------------===//
 
-/// Rewrites 2-D convolution ops with size-1 window dimensions into 1-D
-/// convolution ops. Works with both named ops and equivalent generic ops.
-template <typename Conv2DOp, typename Conv1DOp>
-struct DownscaleSizeOneWindowed2DConvolution final
+/// Rewrite 2-D convolution/pooling/depthwise ops with size-1 window dimensions
+/// into lower-dimensional ops. Uses `inferConvolutionDims` to work with any
+/// layout and handles both named ops and equivalent linalg.generic ops
+/// uniformly. The result is specialized back to a named op when possible.
+FailureOr<LinalgOp> downscaleSizeOneWindowedConvolution(RewriterBase &rewriter,
+                                                        LinalgOp op);
----------------
Abhishek-Varma wrote:

Added a TODO for n-D to (n-1)-D.

https://github.com/llvm/llvm-project/pull/180586