OpenMP and nested parallelism

Question

I would like to "nest" parallel for using OpenMP. Here is a toy code:

#include <iostream>
#include <cmath>

void subproblem(int m) {
  #pragma omp parallel for
  for (int j{0}; j < m; ++j) {
    double sum{0.0};
    for (int k{0}; k < 10000000; ++k) {
      sum += std::cos(static_cast<double>(k));
    }
    #pragma omp critical
    { std::cout << "Sum: " << sum << std::endl; }
  }
}

int main(int argc, const char *argv[]) {
  int n{2};
  int m{8};

  #pragma omp parallel for
  for (int i{0}; i < n; ++i) {
    subproblem(m);
  }

  return 0;
}

Here is what I want:

• If n >= (number of cores on my machine), I want only the first loop to be parallelized.
• If n < (number of cores on my machine), I want OpenMP to launch thread in the inner loop, but I don't want the total number of threads to exceed the number of cores on my machine.

So far, I have only found a solution that disables nested parallelism or always allow it, but I am looking at a way to enable it only if the number of threads launched is below the number of cores.

Is there an OpenMP solution for that using tasks?

Answer 1

Rather than using a pair of nested parallel sections, you can tell OpenMP to "collapse" the nested loops into a single parallel section over the n*m iteration space:

#pragma omp parallel for collapse(2)
for (int i{0}; i < n; ++i) {
  for (int j{0}; j < m; ++j) {
    // ...
  }
}

This will allow it to divide the work appropriately regardless of the relative values of n and m.