RafLit commented on issue #20675:
URL: 
https://github.com/apache/incubator-mxnet/issues/20675#issuecomment-1003948668


   @matteosal 
   I found that the slowdown is caused by enabling large tensor support (the 
-DUSE_INT64_TENSOR_SIZE cmake flag). There are several things that can be done 
to get rid of the slowdown:
   1. Switching from ilp to lp and disabling the large tensor support (if you 
don't need it).
   2. Compiling with onednn enabled - it is a lot faster and doesn't suffer 
from the slowdown.
   3. The individual naive operators used by you, like deconvolution, can be 
looked into and possibly optimized with large tensor support enabled.
   
   I ran some tests with configurations proposed by me and got the following 
results for mxnet 2.0:
   large tensor support enabled, ilp, USE_ONEDNN=OFF: 95 inputs/s
   large tensor support disabled, lp, USE_ONEDNN=OFF: 231 inputs/s
   large tensor support enabled, ilp, USE_ONEDNN=ON: 1050 inputs/s
   large tensor support disabled, lp, USE_ONEDNN=ON: 1154 inputs/s
   
   
   > I mean, do you think that the script should be changed to check the MSVC 
version and fail if it's older than VC2019? I can open a PR adding this check
   
   I think it's a good idea, but it would probably be better to just display a 
warning about a potential slowdown instead of failing.


-- 
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]



---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to