Parallelize recursive function with OpenMP v.2.0

问题

I'm trying to parallelize parts of a project that relies on a lot of recursive algorithms.

Most of them are some form of binary tree creation or traversal and processing.

I'm stuck using GCC v. 4.1.2 on RedHat and the VC++ compiler on Windows (both don't support OpenMP 3.0 with its convenient task construct). I found this question which seems to get the job done with nested parallel sections and some throttling to prevent an exorbitant number of threads.

My question: any way to avoid this approach? Some of the functions are called at each timestep, the overhead of repeatedly creating and destroying a thread team is unacceptable.

Here's the basic structure of a recursive function I've been using, in line with the linked question:

extern int threads;
omp_set_nested(1); omp_set_num_threads(2);
void cell::updateRecursive() {

    // do stuff for cell for this timestep

   #pragma omp flush(threads)
   if (threads>=omp_get_num_procs()) {
       child0->updateRecursive(); child1->updateRecursive(); // no new threads
   } else {
       #pragma omp atomic
       threads++;
       #pragma omp flush(threads)
       #pragma omp parallel sections nowait
       {
           #pragma omp seciton
           child0->updateRecursive();
           #pragma omp section
           child1->updateRecursive();
       }
       #pragma omp atomic
       threads--;
   }
}

This would be sufficient if this function weren't called so often. I'd like a way to recurse that can use an existing team of threads to perform work, rather than creating one while traversing the tree.

Would this be possible at all without task? I've experimented with simply using sections, but apparently they can't be nested to use an existing team of threads.

来源：https://stackoverflow.com/questions/17347252/parallelize-recursive-function-with-openmp-v-2-0

标签

c++

recursion

parallel-processing

openmp