Fix potential deadlock in the WaitEvent path of CmdBuffers #2481
Conversation
1743ecf to
5cd90c9
Compare
Moves the call to event reset for the WaitEvent and AllResetEvent to the ComputeCommandList. This fixes a potential race condition where, if the SignalCommandList executes before the ComputeCommandList, the WaitEvent could be reset before the ComputeCommandList can wait on it and, consequently, create a deadlock.
5cd90c9 to
017d9cc
Compare
I guess that moving the resets to the ComputeCommandList doesn't work when using the CopyCommandList because it might reset the events before the CopyCommandList starts. For that to work we would have to add extra synchronization between the CopyCommandList and the ComputeCommandList. I'm trying a different apprach at the moment with extra barriers in the SignalCommandList. |
There was a problem hiding this comment.
Makes sense to me
I think https://github.com/intel/llvm/blob/sycl/sycl/doc/design/images/L0_UR_command-buffer-v5.jpg could be updated in the DPC++ PR but there's probably a whole bunch of places that doc section isn't quite right, and not sure it's updating giving we're trying to depreciate this path, so leave it up to you whether to do that or not.
Add barriers to the SignalCommandList that guarantee that resetting the WaitEvent is done at the right time.
This fixes a potential race condition where, if the SignalCommandList executes before the ComputeCommandList, the WaitEvent could be reset before the ComputeCommandList can wait on it and, consequently, create a deadlock.