Open josel-amd opened 1 month ago
i dont know if anyone here would be interested, but a few months ago I started working on unifying grouped and depthwise convs by creating an interface for them. At the time I couldn't find many people that cared enough about it so I lost steam with it. I recently resurrected it https://github.com/llvm/llvm-project/pull/94796. This is still a work in progress. I actually first started doing this with depthwise only https://github.com/llvm/llvm-project/pull/75017. This is more complete, but I then decided to switch to grouped conv and make depthwise a specialization of it.
@josel-amd How high priority is 3d grouped convolution for your team?
If we need the support immediately, I could temporarily lower this signature to a linalg generic until a better upstream fix is landed. The longer term convolution changes mentioned by @srcarroll sound attractive in comparison to upstreaming (yet another) convolution named op to the linalg dialect.
@zjgarvey I can't discuss priorities but this is part of a wider effort to get comprehensive support for onnx (you can see that by looking at the other issues I've raised). This is important for us but not urgent. I wonder if some kind of action should still be taken. However, I don't have enough information to dare suggesting a course of action. We can follow the proper way here.
Hi all,
Currently observing a failure when trying to lower the torch.aten.convolution operator. The failure message that I see is:
unimplemented: only 2D grouped convolution supported
.Command to reproduce:
Snippet to reproduce the error is in [0]
Observed result
[0]