Describe the bug
Enabling the version of nonmonotonic dynamic loop scheduling which tries to avoid atomics on the normal, consume local work, path fails on X86.
When trying to run the benchmarks in microBM/scheduling.cc the nonmonotonic benchmarks occasionally assert on Skylake since the "no work available anywhere" test sees that more iterations have been started than exist!
To Reproduce
Disable the #if (1) to change the implementation of
Describe the bug Enabling the version of nonmonotonic dynamic loop scheduling which tries to avoid atomics on the normal, consume local work, path fails on X86. When trying to run the benchmarks in microBM/scheduling.cc the nonmonotonic benchmarks occasionally assert on Skylake since the "no work available anywhere" test sees that more iterations have been started than exist!
To Reproduce Disable the
#if (1)to change the implementation ofin loops.cc to use the implementation which avoids an atomic operation when returning local work.
Run the test code on X86_64.
Desktop (please complete the following information):