c++ - 如何从 OpenMP 单 block 广播结果

Question

我遇到了一个看似微不足道的问题(也许现在还为时过早)。从在 omp parallel 区域内调用的函数中，我想返回一个值，其中 (1) 在每个线程上都必须相同，但 (2) 在 omp single 中计算> 阻止。我怎样才能做到这一点？这是一个小代码片段:

// must work correctly when called from within omp parallel region
// this version is flawed. please suggest corrections
int func(vector<int>&args)
{
  int x;
#pragma omp single
  x = foo(args);    // foo() must be called in single block
                    // would need to broadcast x to all threads here
  return x;         // error: x only set for thread which executed foo()
}

// possible usage (elsewhere, you're not allowed to change this code)
#pragma omp parallel
{
   /* ... */
   int y = func(args);
   /* ... */
}

一个相当不优雅的解决方案是将 single block 转换为 parallel for with reduction:

int func(types args)
{
  int x=0;
#pragma omp for reduction(+:x)
  for(int i=0; i<1; ++i)
    x = foo(args);
  return x;
}

但肯定有更好的解决方案。

Answer 1

解决方案非常简单 - 只需附加 copyprivate(x) 子句，它就是这样做的 - 它广播所列私有(private)变量实例的值single 构造到其他线程:

int func(vector<int>&args)
{
  int x;
#pragma omp single copyprivate(x)
  x = foo(args);    // foo() must be called in single block
                    // would need to broadcast x to all threads here
  return x;
}