#pragma omp并行永不结束

Question

我的代码如下，由于某种原因，我的程序似乎永远不会退出parallel区域。

我的代码获取到

cout<< omp_get_thread_num()<<" done"<< endl; 

线，但它仅有史以来根（0）的过程中执行的，它从不获取到

cout<<"done2?"<< endl; 

一部分。

这就像它已进入无限循环或什么。我甚至没有意识到每个线程都不会在cout之前完成它，因为我在它之前添加了omp barrier，所以每个线程在根甚至可以说完成之前必须达到它。感谢您的帮助。

#pragma omp parallel num_threads(desiredNumThreads) 
{ 
    double startTime; 
    #pragma omp master 
    { 
     startTime = omp_get_wtime(); 
    }   
    while (!Converged) 
    { 
     #pragma omp master 
     { 
      iteration++; 
     } 
     #pragma omp for 
     for (int i = 0; i < matrixsize; i++) 
     { 
      oldX[i] = newX[i]; 
      //newX[i] = Bmatrix[i]; 
     } 
     #pragma omp for 
     for (int i = 0; i < matrixsize; i++) 
     { 
      newX[i] = Bmatrix[i]; //new part 
      for (int j = 0; j < matrixsize; j++) 
      { 
       if (j != i) 
       { 
        newX[i] = newX[i] - matrix[i][j] * oldX[j]; 
       } 
      } 
      newX[i] = newX[i]/matrix[i][i]; //new part 
     } 
     #pragma omp master 
     { 
      Converged = true; 
     } 
     #pragma omp barrier 
     #pragma omp for 
     for (int i = 0; i<matrixsize; i++) 
     { 
      if (abs(oldX[i] - newX[i]) > Tolerance) 
      { 
       Converged = false; 
      } 
     } 
    } 
    #pragma omp master 
    { 
     double endTime = omp_get_wtime(); 
     cout << "This took " << iteration << " iterations. " << endl; 
     cout << "This took " << endTime - startTime << " seconds" << endl; 
     cout << "File to Print answer to? "; 
     cin >> filename; 
     ofstream fout(filename); 
     for (int i = 0; i < matrixsize; i++) 
     { 
      fout << fixed << newX[i] << endl; 
     }     
    } 
    #pragma omp barrier 
    cout<<omp_get_thread_num()<<" done"<<endl; 
} 
cout<<"done2?"<<endl; 

Answer 1

那么我通过移动最后一节

#pragma omp master 
 
    { 
 
     double endTime = omp_get_wtime(); 
 
     cout << "This took " << iteration << " iterations. " << endl; 
 
     cout << "This took " << endTime - startTime << " seconds" << endl; 
 
     cout << "File to Print answer to? "; 
 
     cin >> filename; 
 
     ofstream fout(filename); 
 
     for (int i = 0; i < matrixsize; i++) 
 
     { 
 
      fout << fixed << newX[i] << endl; 
 
     }     
 
    }

出并行sectio的固定它并顺序执行（因为它已经是连续的）。 为什么这解决了我无能为力的问题。

Answer 2

根据MSDN，线程的最大数目是10更改desiredNumThreads至10或更小，它应该工作。

Answer 3

根据注释，代码在Debug配置中工作，但不在Release配置中。这表明编译器正在进行阻止代码运行的优化。这些优化可能来自我建议阅读OpenMP内存模型的OpenMP内存模型）。原始代码中存在一些缺陷（语句中的竞争条件（“Converged = true”））。添加适当的同步将迫使编译器正确地更新变量Converged并可能使问题消失

干杯， -Michael