[GitHub] [incubator-mxnet] apeforest edited a comment on issue #14496: performance degradation from 1.3.1 to 1.4.0

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

apeforest edited a comment on issue #14496: performance degradation from 1.3.1 to 1.4.0
URL: https://github.com/apache/incubator-mxnet/issues/14496#issuecomment-475887608
 
 
   I could verify the performance degradation using Sam's script on the transpose operator.
   
   The performance slow down is mainly due to arithmetic operation performance difference between int32_t and int64_t data types. I changed the data type in Tensor object in mshadow from index_t (which is typedef to int64_t) to int32_t and can see runtime of transpose operator come back to almost same as in 1.3.1. 
   
   I am doing the following experiment: 
   
   1) casting tensor index data type to int32_t at runtime based on tensor size 
   2) using a env variable to select between a large tensor object (defined by int64_t) and small tensor object (defined by int32_t) 
   
   I will post the experiment results and from it see how we can generalize one of these approaches.
   
   @pengzhao-intel Any compiler optimization flags for int64_t data types in Intel architecture? Also, how did MKLDNN handle int64_t performance? Your insight will be appreciated.

----------------------------------------------------------------
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:
users@infra.apache.org


With regards,
Apache Git Services