PhilippvK commented on PR #14272: URL: https://github.com/apache/tvm/pull/14272#issuecomment-1468690703
> One way that I can think of, although it's not very elegant, is to allow a special nn.conv2d_depthwise parameter thorough desired_layout_ops, then override the FTVMConvertOpLayout function for nn.conv2d before running the pass similar to what is done [here](https://github.com/apache/tvm/blob/main/python/tvm/relay/op/contrib/arm_compute_lib.py#L115). Perhaps we can leave as a "TODO" for now and create a tracking issue for it? Thank you for pointing that out. If I understand it correctly, we than need to run the `ConvertLayout` pass twice since we would not be able to pass layouts for both types of convolutions at the same time: 1. With `FTVMConvertOpLayout` being a no-op for **depthwise** convolutions and desired layout for `nn.conv2d` 2. With `FTVMConvertOpLayout` being a no-op for **non-depthwise** convolutions and desired layout for `nn.conv2d` I agree that this is far away from elegant and we should track this somewhere else... -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: commits-unsubscr...@tvm.apache.org For queries about this service, please contact Infrastructure at: us...@infra.apache.org
