CODEWITHSUNDEEP
c++nestedopenmpnested-loops
Akiiino
Akiiino

Reputation: 1100

OpenMP; action between nested loops

So, I have roughly this code:

for (int i = 0; i != 10000; ++i) {
    doAction(i);
    for (int j = 0; j != 10000; ++j) {
        ...
    }
}

And I want to parallellize it using OpenMP. As I understand, a simple collapse won't do in this case; my attempts to use separate #pragma omp fors have borne no fruit either. Is there a simple way to parallelize this easily or do I have to resort to calling doAction i*j times?

Upvotes: 0

Views: 92

Answers (1)

Soonts
Soonts

Reputation: 21956

The simple way to parallelize, only use OpenMP for the outer loop.

Parallelizing stuff all the way down isn’t a good thing, because thread synchronization & task scheduling overhead. When you split a large CPU bound task into pieces for parallel execution, ideally the pieces should be as large as possible while using all available CPU cores most of the time.

P.S. If you have OpenMP 4, for the inner loop, you might want to #pragma omp simd instead of parallel. The outer loop should still be parallel. This way you'll use both kinds of parallelism at the same time, the outer loop parallelized across cores, the inner loop parallelized across SIMD lanes. Theoretically, that's often the fastest way to compute stuff.

Upvotes: 1

Related Questions