Page MenuHomePhabricator

[openmp] Fixed nonmonotonic schedule implementation.

Authored by AndreyChurbanov on Jun 1 2020, 11:27 AM.



Threads may work on different loops concurrently, thus iterations stealing algorithm should check that thief and victim work on the same loop.
The changes are:

  • th_steal_lock moved from thread-private to thread-loop-private buffer;
  • fixed dispatch_init to check possibly corrected in dispatch_init_algorithm schedule;
  • fixed stealing process and stealing finalization to work with thread-loop-private buffers of the same index, instead of th_dispatch_pr_current which might point to different loops in different threads.

Diff Detail

Event Timeline

hbae accepted this revision.Jun 3 2020, 8:43 AM


This revision is now accepted and ready to land.Jun 3 2020, 8:43 AM
This revision was automatically updated to reflect the committed changes.