FrozenGene commented on issue #4803: [WIP][Frontend] Asymmetric padding of convolution support URL: https://github.com/apache/incubator-tvm/pull/4803#issuecomment-592592327 > So, if I understand correctly there are 2 issues here > > * If you try to pass the asymmetric padding via QNN conv2d, QNN conv2d is not able to handle 4 dimensions. I agree that needs to change. > * Currently, even if I make QNN conv2d handle 4 dimensions, QNN calls Pad explicity, which might lead to sub-optimal performance. So, suggestion here is to add a pad_const value in Conv2D attrs and let Conv2D schedule handle padding. Correct? > > I can handle them in 2 separate PRs. For Conv2D attrs, I will have to put RFC to gather everybody thoughts on changing the Conv2D API. Not sure, if everybody will be happy with changing extensively used Conv2D API. > > @FrozenGene do you have any rough numbers for what is the padding overhead if we have it a separate operator? For C5 Cascade/Skylake servers, I don't see pad taking any significant time. So, just curious. (maybe bad schedule for other platforms) ping @anijain2305 have you plan to handle it recently?
---------------------------------------------------------------- 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. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services
