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

[Abhishek Varma]

Abhishek-Varma wrote:

> I haven't reviewed the details yet because we need to be aligned on the pass definition. Please take a look at inlined comments.
> 
> btw, this does not support batch-less conv right? We should do it in a follow-up (not this PR) if possible.

Hi @hanhanW - I've updated the pass to now specialize when possible. This really helped get rid of the workaround in lit tests and made it super clean! Thanks for the comment.

And this does support batchless convs - I've added a lit test for each of [conv/depthwise_conv/pool] variants now to demonstrate the same.

Have updated the PR description too.

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