llvm
llvm copied to clipboard
[SYCL] optimize wait() for in-order queue
trafficstars
This shortens the critical section for in-order and out-of-order cases and avoids taking the lock entirely is MNoEventMode is set.
Performance results are a bit noisy, but it seems there is some improvement on BMG: https://oneapi-src.github.io/unified-runtime/performance/?runs=Baseline_PVC_L0v2%2CBaseline_BMG_L0v2%2Cv2_wait_opt_pvc%2Cv2_wait_opt_bmg
@sergey-semenov - can you review this?