[GitHub] zheng-da commented on issue #10215: Avoid MKLDNN in-place data conversion in NDArray synchronously

[GitBox <gi...@apache.org>]

zheng-da commented on issue #10215: Avoid MKLDNN in-place data conversion in NDArray synchronously
URL: https://github.com/apache/incubator-mxnet/pull/10215#issuecomment-375844272
 
 
   @TaoLv It's difficult. We discussed this before. Who should add this "reorder" operator? Model developer? It won't be user-friendly.
   
   MXNet itself? To add the reorder in the right place, we need to know both operators. An operator only knows itself. So NNVM? If so, we need to make NNVM aware of MKLDNN. It's supposed to be a very general framework. I don't think it's a good idea to do it.
   
   I have code that converts data layout in the executor (not adding reorder as an operator), so operator developers don't need to convert layout themselves. We just don't have the right test case for the code. I'll create another PR for the layout conversion once the bug in sparse operators is fixed.

----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
 
For queries about this service, please contact Infrastructure at:
users@infra.apache.org


With regards,
Apache Git Services