search

zy97140 / omp-benchmark-for-pytorch

Benchmark omp threshold for pytorch.
5 stars 2 forks source link

How to test performance of different openmp theshold? #1

Closed XiaobingSuper closed 6 years ago

XiaobingSuper commented 6 years ago

I want to test the performance by setting a different omp theshold, I changed the OMP_OVERHEAD_THRESHOL numbers in [THTensorApply.hpp]((https://github.com/pytorch/pytorch/blob/master/aten/src/TH/generic/THTensorApply.hpp) for series or parallel and made a test in Platinum 8180 , but I found there has a little difference between seris and parallel. Can you tell me how do you set the omp theshold number to get you results? Thanks!

MlWoo commented 6 years ago

@zxb1489479870 At first, make sense the following rules:

  1. all the operations will be executed in series by setting the Threshold Macro to a very, very large value(eg. 1000k)
  2. all the operations will be executed in parallel by setting the Threshold Macro to a very, very tiny value(eg. 500)

Then check the perf boost radio between parallelization and seriation by varying the size of tensor. The threshold will be determined according to the boost status. Our optimization has been merge into the master branch. you'd better check the code in my pr.