Skip to content

WIP: Replace GpuEventSynchronizer with DeviceEventSynchronizer

Andrey Alekseenko requested to merge 2527-gpueventsynchronizer-improvements into 3924-converge-gpueventsynchronizer-and-commandevent

Major differences:

  • markEvent and enqueueWaitEvent functionality moved to DeviceStream class.
  • DeviceEventSynchronizer is backend-agnostic, no more different implementations for CUDA, OpenCL, and SYCL.
  • More configurable consumption counting. Previously, we had 1:1 on OpenCL and SYCL, and no counting on CUDA.
  • Added a check whether the event was marked.

Notes:

  • The new counting feature has not been properly enabled yet to avoid breaking CUDA code.
  • Event marking is not used.

These limitations will be addressed in follow-up MRs. Here we aim to preserve the existing logic.

Overall, the idea is that we can do any synchronization we want with DeviceEvent and DeviceStream, while DeviceEventSynchronizer serves as a convenience wrapper with additional producer-consumer accounting.

We might also wish to replace some uses of GpuEventSynchronizer with a simple DeviceEvent when no accounting is needed.

Closes #2527

Edited by Andrey Alekseenko

Merge request reports