C++ Armadillo 和 OpenMp : Parallelization of summation of outer products - define reduction for Armadillo matrix

Question

我正在尝试使用 OpenMP 并行化一个 for 循环，它对 Armadillo 矩阵求和。我有以下代码:

#include <armadillo>
#include <omp.h>

int main()
{

        arma::mat A = arma::randu<arma::mat>(1000,700);
        arma::mat X = arma::zeros(700,700);
        arma::rowvec point = A.row(0);

        # pragma omp parallel for shared(A) reduction(+:X)
        for(unsigned int i = 0; i < A.n_rows; i++){
                arma::rowvec diff = point - A.row(i);
                X += diff.t() * diff; // Adding the matrices to X here
        }

}

我收到这个错误:

[Legendre@localhost ~]$ g++ test2.cpp -o test2 -O2 -larmadillo -fopenmp
test2.cpp: In function ‘int main()’:
test2.cpp:11:52: error: user defined reduction not found for ‘X’

我阅读了有关定义缩减的内容，但我还没有找到使用 Armadillo 矩阵的示例。就我而言，定义 Armadillo 矩阵缩减的最佳方法是什么？

Answer 1

这些缩减仅适用于内置类型(double、int 等)。因此，您必须自己进行还原，这很简单。只需将每个线程的结果累积到一个线程局部变量中，然后将其添加到临界区内的全局结果。

#include <armadillo>
#include <omp.h>

int main()
{

  arma::mat A = arma::randu<arma::mat>(1000,700);
  arma::mat X = arma::zeros(700,700);
  arma::rowvec point = A.row(0);

  #pragma omp parallel shared(A)
  {
    arma::mat X_local = arma::zeros(700,700);

    #pragma omp for
    for(unsigned int i = 0; i < A.n_rows; i++)
    {
      arma::rowvec diff = point - A.row(i);
      X_local += diff.t() * diff; // Adding the matrices to X here
    }

    #pragma omp critical
    X += X_local;
  }
}

使用更新的 OpenMP(我认为是 4.5？)，您还可以为您的类型声明用户定义的缩减。

#include <armadillo>
#include <omp.h>

#pragma omp declare reduction( + : arma::mat : omp_out += omp_in ) \
  initializer( omp_priv = omp_orig )

int main()
{

  arma::mat A = arma::randu<arma::mat>(1000,700);
  arma::mat X = arma::zeros(700,700);
  arma::rowvec point = A.row(0);

  #pragma omp parallel shared(A) reduction(+:X)
  for(unsigned int i = 0; i < A.n_rows; i++)
  {
    arma::rowvec diff = point - A.row(i);
    X += diff.t() * diff; // Adding the matrices to X here
  }
}