emfomenk commented on issue #17980: When compiled with MKL, fully_connected 
calls DNNL while dot and batch_dot call MKL
URL: 
https://github.com/apache/incubator-mxnet/issues/17980#issuecomment-610441396
 
 
   > Do you think that enabling this option can improve the performance of 
inner product primitive for the above cases?
   
   @TaoLv, yes, this could be. Two reasons for that:
   1. If Intel MKL is linked, then most likely `libiomp5.so` will be used 
instead of `libgomp.so`. So, if you see the performance difference, I would 
suggest to test Intel MKL-DNN v1.0 / DNNL v1.1+ with `LD_PRELOAD=libiomp5.so`, 
just to confirm that the issue due to or not due to OpenMP run-time.
   2. Performance bug in Intel MKL-DNN v1.0 / DNNL v1.1+. We definitely fixed 
few performance issues in the library since v1.0.
   
   I would also suggest to try upgrading the library to the most recent version 
and check the performance.
   I strongly suggest not using this developer's option. It is unsupported, not 
tested, and as I said, it is most likely broken in the subsequent releases.

----------------------------------------------------------------
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

Reply via email to