omp/reduce.h

Go to the documentation of this file.

#ifndef GRPPI_OMP_REDUCE_H
#define GRPPI_OMP_REDUCE_H

#ifdef GRPPI_OMP

#include "parallel_execution_omp.h"
#include "reduce.h"

namespace grppi {

template < typename InputIt, typename Identity, typename Combiner>
auto reduce(parallel_execution_omp & ex, 
            InputIt first, InputIt last, 
            Identity identity,
            Combiner && combine_op)
{
    int numElements = last - first;
    int elemperthr = numElements/ex.concurrency_degree();
    auto identityVal = identity;
    //local output
    std::vector<typename std::iterator_traits<InputIt>::value_type> out(ex.concurrency_degree());
    //Create threads
    #pragma omp parallel
    {
    #pragma omp single nowait
    {
    for(int i=1;i<ex.concurrency_degree();i++){

      auto begin = first + (elemperthr * i);
      auto end = first + (elemperthr * (i+1));
      if(i == ex.concurrency_degree() -1) end = last;

         #pragma omp task firstprivate (begin, end,i)
         {
            out[i] = identityVal;
            while( begin != end ) {
                out[i] = combine_op(*begin, out[i] );
                begin++;
            }
         }

    }
    
    //Main thread
    auto end = first + elemperthr;
    out[0] = identityVal;
    while(first!=end){
         out[0] = combine_op(*first , out[0]);
         first++;
    }

    #pragma omp taskwait
    }
    }

    auto outVal = out[0];
    for(unsigned int i = 1; i < out.size(); i++){
       outVal = combine_op(outVal, out[i]);
    }
    return outVal;
}

}



#endif

#endif