Skip to content

Implement better error propagation and semaphore

Churavy, Valentin requested to merge vc/events_take_two into master

Created by: vchuravy

@lcw observed that this delightful code deadlocks:

    barrier = Base.Threads.Event()
    cpu_event = Event(wait, barrier)

    wait(CUDA(), cpu_event) # Event edge on CuDefaultStream
    gpu_event = Event(CUDA()) # Event on CuDefaultStream

    notify(barrier)
    wait(gpu_event)

The reason was that I thought it would be okay wait(CUDA(), cpu_event) # Event edge on CuDefaultStream to be a blocking event. not thinking about the fact that there could be dependency edges outside our control (here the barrier) in Lucas original code MPI

Merge request reports