c++ logging from parallelized loop

Question

I have a parallel for loop and ONCE for looping I'd like to log some details.

first_iteration = true;

#pragma omp parallel for schedule(runtime) 
for (int j = 0; j < items.size(); j++) {

    if (first_iteration) {
        FILE_LOG(logINFO) << "blablabla";
        first_iteration = false;
    }

    // do something
}

But it is always logged several times (probably number of thread times). And I have a suspicion that time to time it throws a segfault. I've read a possible solution with temp variable and #pragma omp atomic read and write. But it doesn't seem to work with Visual Studio 19 (c++std14):

error C3005: 'read': unexpected token encountered on OpenMP 'atomic' directive`.

I need it work also in g++ 11.3.0.

Answer 1

If you want to stick with OpenMP atomics, try this. It uses a test and test-and-set logic to avoid cache-line bouncing and needless synchronization overhead.

int main()
{
    int iteration = 0;

#   pragma omp parallel for
    for(int j = 0; j < 100; j++) {
        int own_iteration;
#       pragma omp atomic read relaxed
        own_iteration = iteration;
        if(own_iteration == 0) {
#           pragma omp atomic update relaxed capture
            own_iteration = iteration += 1;
            if(own_iteration == 1)
                std::cout << "blablabla\n";
        }
    }
}

You can make the code simpler with C++-11 atomics.

#include <atomic>

int main()
{
    std::atomic<bool> first_iteration = true;

#   pragma omp parallel for
    for(int j = 0; j < 100; j++) {
        if(first_iteration.load(std::memory_order_relaxed)
                && first_iteration.exchange(false, std::memory_order_relaxed)) {
            std::cout << "blablabla\n";
        }
    }
}

Of course "first iteration" doesn't really mean much in fully parallel execution. You might as well check if(j == 0) and be done with it.