parallel_for_

parallel

asked 2013-08-15 03:28:51 -0600

updated 2013-08-15 03:29:47 -0600

I'm not sure, that I can understand parallel_for_... I wrote simple class:

//header
class ParallelClass: public cv::ParallelLoopBody
{
    private:

    public:
    ParallelClass();
    void operator()( const cv::Range &r )const;
};
//cpp
 ParallelClass::ParallelClass(){}
 void ParallelClass::operator()( const cv::Range &r )const{
    cout<<"start"<<endl;
  }

and in the main function:

  parallel_for_(cv::Range(0,10), ParallelClass());

I was expecting to see ten times "start",,, but looks like I'm totally wrong...

edit retag flag offensive close merge delete

Comments

Works perfectly fine for me. Have you forgot to build with tbb-support? Also note (not directly related to your problem) that you should later write sth like for (int i = r.start; i < r.end; ++i){} to give each thread some more to do.

Guanta ( 2013-08-15 03:54:19 -0600 )edit

add a comment

3 answers

Sort by » oldest newest most voted

answered 2013-08-15 18:36:29 -0600

Mathieu Barnachon

4678 ●18 ●53 http://www.math-barnac...

updated 2013-08-18 17:06:53 -0600

If you don't see ten times "start" it's because the parallel framework doesn't launch as many threads as your range. Write this instead:

void ParallelClass::operator()( const cv::Range & r ) const {
    // loop for the range
    for( int i = r.start ; i < r.end ; ++r ) {
        std::cout << "start" << std::endl;
    }
}

Therefore you will see your start written 10 times. The reason is the parallel binding used (concurency, tbb, etc.) cut your loop in smaller pieces, and you still need to execute your code for a range (argument is a range not an index), sometimes the range could be one, but it's an exception, not the norm.

edit flag offensive delete link

Comments

Shouldn't he have seen 10 times "start" even without the for-loop? At least it works perfectly fine for me without it, however, I guess that only one thread actually executes this part then...

Guanta ( 2013-08-16 03:44:52 -0600 )edit

It depends on your hardware (and probably the binding used: OpenMP, TBB, Concurrency, c=, ...). You should print the index (r.start and r.end) to see what happens on your system.

Mathieu Barnachon ( 2013-08-16 18:10:17 -0600 )edit

@Mathieu, there a typo in your 'for' loop. It should be: for( int i = r.start ; i < r.end ; ++i )

Michael Burdinov ( 2013-08-18 06:34:25 -0600 )edit

Thanks, I fix it!

Mathieu Barnachon ( 2013-08-18 17:09:08 -0600 )edit

add a comment

answered 2013-08-18 06:24:51 -0600

Michael Burdinov
4638 ●6 ●33 ●86

When I encountered parallel_for_ for the first time I posted here a question about how to use it properly. I got very good answers from @alexcv and @Vladislav Vinogradov. I know that this is not a whole answer to your question, i.e. it doesn't explain why your approach is not working, but at least it shows you what approach will work.