← Back to events
ActiveAIRelease3.5

viable/strict/1788292262: [ROCm] Set USE_ARC in ROCm workflows (#194123)

Photo: PyTorch Releases

What happened

The USE_ARC variable will be used to export OMP_NUM_THREADS in test.sh . According to the script, setting OMP_NUM_THREADS to a number less than the number of available cores can avoid OpenMP's busy waiting, therefore improve performance. This parity report shows a speedup of ~3.5% compared to main trunk. It's a comparison between this PR's trunk run and a recent trunk run on main TOTAL 04A989B1 RUNNING TIME 104578.7…

Summary assembled by rule from the sources below

Why it's spreading

Sources