如何在MPI + openmp中啟動多線程？

Question

我想在我的MPI應用程序代碼中的一個進程中啟動OpenMP多線程區域。 例如：

#include <iostream>
#include <omp.h>
#include <mpi.h>
#include <Eigen/Dense>
using std::cin;
using std::cout;
using std::endl;

using namespace Eigen;

int main(int argc, char ** argv)
{
    int rank, num_process;
    MatrixXd A = MatrixXd::Ones(8, 4);
    MatrixXd B = MatrixXd::Zero(8, 4);
    MPI_Init(&argc, &argv);
    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
    MPI_Comm_size(MPI_COMM_WORLD, &num_process);
    MPI_Status status;
    if (rank == 0)
    {
        int i, j, bnum = 2, brow = 4, thid;
        #pragma omp parallel shared(A, B) private(i, j, brow, bnum, thid) num_threads(2)
        for (i = 0; i < brow; i ++)
        {
            for (j = 0; j < 4; j ++)
            {
                thid = omp_get_thread_num(); 
                //cout << "thid " << thid << endl;
                B(thid * brow+i,j) = A(thid*brow+i, j);
            }
        }
        cout << "IN rank 0" << endl;
        cout << B << endl;
        cout << "IN rank 0" << endl;
        MPI_Send(B.data(), 32, MPI_DOUBLE, 1, 1, MPI_COMM_WORLD);
    }
    else
    {
        MPI_Recv(B.data(), 32, MPI_DOUBLE, 0, 1, MPI_COMM_WORLD, &status);
        cout << "IN rank 1" << endl;
        cout << B << endl;
        cout << "IN rank 1" << endl;
    }
    MPI_Finalize();
    return 0;
}

在我的示例代碼中，我想啟動2個線程將數據從矩陣A復制到矩陣B，並且我的機器具有4個內核。 但是，當運行程序時，矩陣B僅獲得一半的數據。

$ mpirun -n 2 ./shareMem
IN rank 0
0 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
IN rank 0
IN rank 1
0 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
IN rank 1

$ mpirun -n 4 ./shareMem # it just hang on and doesn't exit
IN rank 0
0 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
IN rank 0
IN rank 1
0 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
IN rank 1

我期望的輸出是

$ mpirun -n 2 ./shareMem # it just hang on and doesn't exit
IN rank 0
1 1 1 1
1 1 1 1
1 1 1 1
1 1 1 1
1 1 1 1
1 1 1 1
1 1 1 1
1 1 1 1
IN rank 0
IN rank 1
1 1 1 1
1 1 1 1
1 1 1 1
1 1 1 1
1 1 1 1
1 1 1 1
1 1 1 1
1 1 1 1
IN rank 1

如何解決該問題並在代碼中運行2個線程？ 謝謝！

Answer 1

並行一詞中有一個錯字，編譯器無法捕獲。

#pragma omp prallel

PS：我沒有足夠的聲譽來添加評論

Answer 2

更改

#pragma omp parallel shared(A, B) private(i, j, brow, bnum, thid) num_threads(2)

至

#pragma omp parallel shared(A, B) private(i, j, thid) num_threads(2)

brow和bnum是共享變量。 通過將名稱bnum和brow添加到private子句，您將為每個線程使用這樣的名稱創建新的自動變量，並且默認情況下它們是未定義的。