You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@mxnet.apache.org by GitBox <gi...@apache.org> on 2018/02/06 19:44:01 UTC

[GitHub] Maratyszcza commented on issue #2986: [call for contribution] Improving CPU performance

Maratyszcza commented on issue #2986: [call for contribution] Improving CPU performance
URL: https://github.com/apache/incubator-mxnet/issues/2986#issuecomment-363541857
 
 
   @xmchen1987 @hjk41 @tqchen I looked at NNPACK bindings in MXNet, and they have room for improvement:
   - NNPACK now includes CMake configuration scripts for all platforms. It is better to use those rather than stick to an old NNPACK version, as NNPACK is getting updates and performance improvements.
   - NNPACK supports using pre-allocated workspace buffers provided by the framework rather then allocating and de-allocating them inside NNPACK on each convolution call. This is a big cost, especially for small convolutions. See Maratyszcza/NNPACK#75 for details.
   - NNPACK supports pre-computing transformed coefficients for inference use-cases (when weights do not change between forward runs). See Maratyszcza/NNPACK#82 for details.
   - NNPACK can do fused Convolution+ReLU at the cost a single convolution operation. See `activation` parameter.

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