Relax cudnn conditions for channels-last convolutions (#38904)
Summary:
Follow up of https://github.com/pytorch/pytorch/issues/38044. Thanks ptrblck, mcarilli for the help on discussing the changes!
Could fix https://github.com/pytorch/pytorch/issues/37725 by skipping the depthwise-workload check introduced in https://github.com/pytorch/pytorch/issues/22302. This PR also relaxed dilated convolution for channels-last.
The testing script is https://gist.github.com/xwang233/82a707f69bb710cb612349280a2c5f41. About 387k conv arguments were tested and no cudnn exception was thrown.
cc ngimel VitalyFedyunin ptrblck mcarilli
Pull Request resolved: https://github.com/pytorch/pytorch/pull/38904
Differential Revision: D22155797
Pulled By: VitalyFedyunin
fbshipit-source-id: 81b5736cec67ea263029121521c6acafd9dddba6
