函数依赖 openMP

Question

我有 5 个函数 A、B、C、D、E。

要执行 D，我需要执行 B、C， 要执行 E 我需要执行 A、D。

我试过了

int main()
{
    #pragma omp parallel num_threads(5) 
    {
        long t1 = clock();
        int a = 0, b = 0, c = 0, d = 0, e = 0;
        int th = omp_get_thread_num();
        if (th == 0) {
            a += A();
            printf("A is finished after %d\n", clock() - t1);
        }
        if (th == 1) {
            b += B();
            printf("B is finished after %d\n", clock() - t1);
        }
        if (th == 2) {
            c += C();
            printf("C is finished after %d\n", clock() - t1);
        }
        if (th == 3 && (b == 1 && c == 1)) {
            d += D();
            printf("D is finished after %d\n", clock() - t1);
        }
        if (th == 4 && (a == 1 && d == 1)) {
            e += E();
            printf("E is finished after %d\n", clock() - t1);
        }

    }
    return 0;
}

但是D、E还没有执行

到目前为止，所有这些函数都返回 1 用于调试目的

Answer 1

变量a、b、c 和d, 不能用于线程间通信，因为它们都是私有。因此，每个线程都有自己的私有副本。此外，将它们用于同步目的通常不是一个好主意。

在您的代码中，thread=3 永远不会等待 if (th == 3 && (b == 1 && c == 1))，因为：

1. b 和 c 是私有的，所以 thread=3 具有 b=0 和 c=0 而不管其他线程对变量 b=0 和 c=0 的 副本 做了什么.
2. 什么都没有 告诉那个线程要等待（例如，一些类似同步的构造函数）。

如果您希望线程相互等待，请改用omp barrier。所有线程都必须调用屏障才能继续计算。

 int main()
    {
        #pragma omp parallel num_threads(5) 
        {
            long t1 = clock();
            int a = 0, b = 0, c = 0, d = 0, e = 0;
            int th = omp_get_thread_num();
            if (th == 0) {
                a += A();
                printf("A is finished after %d\n", clock() - t1);
            }
            if (th == 1) {
                b += B();
                printf("B is finished after %d\n", clock() - t1);
            }
            if (th == 2) {
                c += C();
                printf("C is finished after %d\n", clock() - t1);
            }
            // Threads will wait for each other here
            #pragma omp barrier 
            if (th == 3) {
                d += D();
                printf("D is finished after %d\n", clock() - t1);
            }
            // Threads will wait for each other here
            #pragma omp barrier 
            if (th == 4) {
                e += E();
                printf("E is finished after %d\n", clock() - t1);
            }
        }
        return 0;
    }

更复杂的方法是使用依赖于 OpenMP 4.0 标准上发布的功能的任务。关于这个功能在this thread.上的工作原理已经有了很好的解释

int a = 0, b = 0, c = 0, d = 0, e = 0;
#pragma omp parallel num_threads(5) shared(a, b, c, d)
{
  #pragma omp single nowait
  {
      long t1 = clock();

      int th = omp_get_thread_num();
      #pragma omp task  depend (out:a) 
      {
          a += A();
          printf("A is finished after %d\n", clock() - t1);
      }
      #pragma omp task depend (out:b) 
      {
         b += B();
         printf("B is finished after %d\n", clock() - t1);
      }
      #pragma omp task depend (out:c) 
      { 
          c += C();
          printf("C is finished after %d\n", clock() - t1);
      }
     #pragma omp task depend (in:a, b) depend(out:d) 
     {
        d += D(); 
        printf("D is finished after %d\n", clock() - t1);
     }
     #pragma omp task depend (in:a, b)  
     {
       e += E();
       printf("E is finished after %d\n", clock() - t1);
     }
  }
}
return 0;
}

Answer 2

适当的 OpenMP 解决方案是使用具有数据依赖性的任务：

    #pragma omp parallel num_threads(3)
    #pragma omp single
    {
        double t1 = omp_wtime();
        int a = 0, b = 0, c = 0, d = 0, e = 0;
        #pragma omp task shared(a) depend(out: a)
        {
            a += A();
            printf("A is finished after %f\n", omp_wtime() - t1);
        }
        #pragma omp task shared(b) depend(out: b)
        {
            b += B();
            printf("B is finished after %f\n", omp_wtime() - t1);
        }
        #pragma omp task shared(c) depend(out: c)
        {
            c += C();
            printf("C is finished after %f\n", omp_wtime() - t1);
        }
        #pragma omp task shared(b,c,d) depend(in: b,c) depend(out: d)
        {
            d += D();
            printf("D is finished after %f\n", omp_wtime() - t1);
        }
        #pragma omp task shared(a,d,e) depend(in: a,d)
        {
            e += E();
            printf("E is finished after %f\n", omp_wtime() - t1);
        }

    }

在这里，任务A 被标记为a 的值的生产者depend(out: a)，任务D 被标记为d 的生产者depend(out: d)。任务E 用depend(in: a,d) 标记为这两个值的消费者。在输出（生产者）和输入（消费者）依赖之后，OpenMP 运行时构建一个执行 DAG（有向无环图），告诉它所有任务的正确执行顺序。您也不需要五个线程 - 三个就足够了。

在single 结构中包含任务代码是非常惯用的 OpenMP。

OpenMP 4.0 早在 2013 年就引入了任务依赖项，因此除 MSVC++ 之外的任何现代编译器都应提供对该功能的支持。