Hello,
I'm trying to work out the compute times for every operator in my computation graph when I do a forward pass. Is it possible to obtain this information for GPU and CPU modes? If not, is it possible I can instrument the code in some way to retrieve this information?
Thanks,
Piyush3dB
we currently don't have a profiler. You can try adding it in graph_executor.
Related to https://github.com/dmlc/mxnet/issues/406, tianqi has given some helpful instructions. I was not familiar with MXNet at that time so I had not continued.
ok, thanks for the pointer. @sxjscience, I'm not too familiar with that part of the MXNet code. I'll have a look and work out what to do. Did you have a better idea about it now? Thanks.
@Piyush3dB We can possibly insert timer in the execution logic. I need more tests to confirm the idea. Also, timing may not be trivial due to the asynchronous executions.
Also timing of operations executed on the gpu will need special consideration
Any news of this feature? Is it merged into the master branch?
see examples/profiler
Most helpful comment
we currently don't have a profiler. You can try adding it in graph_executor.