OpenMP thread communication for device operations

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi,
Here’s the scenario:
1st thread performs some device operations and then calls cudaEventRecord(event); (NOTE: This event was created with flag ‘cudaEventBlockingSync’)
2nd thread waits for this event with cudaEventSynchronize(event) and then performs some CPU operations.

Code outline:

for(int i=0;i<5;i++)
    cudaEventCreateWithFlags(&event[i], cudaEventBlockingSync);
omp_set_num_threads(2);
#pragma omp parallel shared(event)
{
    int ti = omp_get_thread_num();

    if(ti == 0)
    {
        for(int i=0;i<5;i++)
        {
            //Some device operations
            cudaEventRecord(event[i]);
        }
    }
    else
    {
        for(int i=0;i<5;i++)
        {
            cudaEventSynchronize(event[i]);    //Doesn't wait till completion of event[i] :(
            //Some CPU operations
        }
    }
}

As shown in the code, the cudaEventSynchronize(event[i]) doesn’t block the CPU thread 1 till completion of device operations done by CPU thread 0 for ‘i’ (As seen in Visual Profiler).

Can anybody tell what I might be doing wrong?
Thanks in Advance! :)

2

I don’t know too much about cudaEvents, but, what if the cudaEventSynchronize(event[i]) executes before cudaEventRecord(event[i])? what should be the behavior?

3

Oh, good point! It will return cudaSuccess and execution will continue!
Thanks for that. I’m thinking of a work-around for this, if you have any suggestions, please reply! :)