CODEWITHSUNDEEP
c++11openmp
Jean-Michaël Celerier
Jean-Michaël Celerier

Reputation: 7886

Using OpenMP with C++11 range-based for loops?

Is there any counter-indication to doing this ? Or is the behavior well specified?

#pragma omp parallel for
for(auto x : stl_container)
{
   ...
}

Because it seems that OpenMP specification is only valid for c++98 but I guess there might be more incompatibilities due to C++11 threads, which are not used here. I wanted to be sure, still.

Upvotes: 91

Views: 26110

Answers (2)

Jean-Michaël Celerier
Jean-Michaël Celerier

Reputation: 7886

OpenMP 5.0 adds the following line on page 99, which makes a lot of range-based for loops OK !

2.12.1.3 A range-based for loop with random access iterator has a canonical loop form.

Source : https://www.openmp.org/wp-content/uploads/OpenMP-API-Specification-5.0.pdf

Upvotes: 37

Hristo Iliev
Hristo Iliev

Reputation: 74435

The OpenMP 4.0 specification was finalised and published several days ago here. It still mandates that parallel loops should be in the canonical form (§2.6, p.51):

for (init-expr; test-expr; incr-expr) structured-block

The standard allows for containers that provide random-access iterators to be used in all of the expressions, e.g.:

#pragma omp parallel for
for (it = v.begin(); it < v.end(); it++)
{
   ...
}

If you still insist on using the C++11 syntactic sugar, and if it takes a (comparatively) lot of time to process each element of stl_container, then you could use the single-producer tasking pattern:

#pragma omp parallel
{
   #pragma omp single
   {
      for (auto x : stl_container)
      {
         #pragma omp task
         {
            // Do something with x, e.g.
            compute(x);
         }
      }
   }
}

Tasking induces certain overhead so it would make no sense to use this pattern if compute(x); takes very little time to complete.

Upvotes: 57

Related Questions