0

I tried to call an OpenBlas function from an OpenMP thread while the Blas parallelisation is set to a value unequal to one. I am using OpenBlas 0.3.9, after downloading the source I untared it and called

make USE_OPENMP=1

make PREFIX=/someFolder/ install

However I always get the following error message from my executeable

OpenBLAS Warning : Detect OpenMP Loop and this application may hang. Please rebuild the library with USE_OPENMP=1 option.

Does anyone know, why this is the case and how I can change it? Here is a minimal example of my code:

#include <complex>
#include <vector>
#include <random>
#include <iostream>
#include <algorithm>
#include <omp.h>
#include <cblas.h>
#include <lapacke.h>

int main(int, char**) {
    int const blas_threads = 2,
              omp_threads = 2,
              matrix_size = 100;
    openblas_set_num_threads(blas_threads);
    omp_set_max_active_levels(2);

    double alpha = 1.,
           beta = 0.;

    std::vector<std::vector<double>> as(omp_threads,
                                        std::vector<double>(matrix_size*matrix_size));
    std::vector<std::vector<double>> bs(omp_threads,
                                        std::vector<double>(matrix_size*matrix_size));
    std::vector<std::vector<double>> cs(omp_threads,
                                        std::vector<double>(matrix_size*matrix_size));

    std::random_device rd;
    std::mt19937 gen(rd());
    std::uniform_real_distribution<double> dis;
    for(int i = 0; i < omp_threads; ++i) {
        std::generate(as[i].begin(),
                      as[i].end(), 
                      [&dis,&gen]() { return dis(gen); });
        std::generate(bs[i].begin(),
                      bs[i].end(), 
                      [&dis,&gen]() { return dis(gen); });
    }

    // for(int i = 0; i < matrix_size*matrix_size; ++i) {
    //  std::cout << as[0][i] << " " << bs[0][i] << std::endl;
    // } 

#pragma omp parallel for num_threads(omp_threads), schedule(static, 1)
    for(int i = 0; i < omp_threads; ++i) {
        cblas_dgemm(CblasColMajor,
                    CblasNoTrans,
                    CblasNoTrans,
                    matrix_size,
                    matrix_size,
                    matrix_size,
                    alpha,
                    as[i].data(),
                    matrix_size,
                    bs[i].data(),
                    matrix_size,
                    beta,
                    cs[i].data(),
                    matrix_size);
    }

    // for(int i = 0; i < matrix_size*matrix_size; ++i) {
    //  std::cout << cs[0][i] << std::endl;
    // } 

    return 0;
}
sam.11100
  • 51
  • 4
  • OpenBlas use already OpenMP in `cblas_dgemm`. Using `#pragma omp parallel` in a loop that call a function already use a `#pragma omp parallel` can cause a significant slowdown and more generally possible issues like this due to the nested `parallel for`. See [here](https://stackoverflow.com/questions/4317551/openmp-what-is-the-benefit-of-nesting-parallelizations) for more information about the issues related to nested parallelism in OpenMP. – Jérôme Richard Mar 09 '20 at 11:23
  • Even though I appreciate your concerns regarding my code quality, that was not the question I asked. Apparently it is possible to do this, since a flag for doing so exists. – sam.11100 Mar 09 '20 at 16:08

0 Answers0