Skip to content
This repository has been archived by the owner on Nov 17, 2023. It is now read-only.

Omp fork numthreads fix 1.6 #17000

Merged
merged 4 commits into from
Dec 12, 2019
Merged

Omp fork numthreads fix 1.6 #17000

merged 4 commits into from
Dec 12, 2019

Commits on Nov 16, 2019

  1. Fix test_gluon.py:test_sync_batchnorm when number of GPUS > 4

    @larroy
    larroy committed Nov 16, 2019
    Configuration menu
    Copy the full SHA
    7fc6255 View commit details
    Browse the repository at this point in the history

Commits on Dec 7, 2019

  1. Prevent after-fork number of OMP threads being bigger than 1. 

    This could happen if it was set in the environment. As we are setting engine::OpenMP::Get()->set_enabled(false) in initialize.cc in the child after forking, the behaviour goes back to what it was before apache#15762 was introduced.

Regions using omp get the threads count from GetRecommendedOMPThreadCount, so if omp is disabled they will get 1 thread and run serially
    @larroy
    larroy committed Dec 7, 2019
    Configuration menu
    Copy the full SHA
    80a82b6 View commit details
    Browse the repository at this point in the history

Commits on Dec 11, 2019

  1. add C++ unit test

    @larroy
    larroy committed Dec 11, 2019
    Configuration menu
    Copy the full SHA
    5b230e0 View commit details
    Browse the repository at this point in the history

  2. Add comment

    @larroy
    larroy committed Dec 11, 2019
    Configuration menu
    Copy the full SHA
    ffd7149 View commit details
    Browse the repository at this point in the history