lhutton1 commented on PR #14272: URL: https://github.com/apache/tvm/pull/14272#issuecomment-1469853396
Thanks for the updates, LGTM after fixing the test failures :) I believe the `ConvertLayout` pass would still only need to be applied once. The `FTVMConvertOpLayout` function can be run on each `nn.conv2d` the `ConvertLayout` pass encounters, check whether the current convolution is normal or depthwise, then use the desired layout specified for `nn.conv2d_depthwise` if the convolution is depthwise and the layout for `nn.conv2d_depthwise` has been specified. Before passing the `desired_layouts` dictionary to the `ConvertLayout` pass itself, we can remove the special `nn.conv2d_depthwise` value. Apologies for the rushed explanation, happy to help explain further in a separate conversation -- 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: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
