multiple reduction on stl::vector gets wrong results

Sean_D_ — Tue, 02 May 2017 17:42:49 GMT

Hi,

In the following test code, I use omp declare reduction on stl::vector. Using icpc 17.0, omp parallel reduction works correctly on a single stl::vector. However, when I attempt to reduce to two separate vectors, v1 & v2, it returns v1=sum(v1)+sum(v2) and v2=0.0 (where sum is across threads). By my reading of the standard, I would expect it to return v1=sum(v1) and v2=sum(v2). This is what GCC 6.3 does.

Can my omp directives be tweaked to work with icpc 17.0? Or is there another problem?

Thanks in advance,

Sean

// http://stackoverflow.com/questions/43168661/openmp-and-reduction-on-stdvector
//
#include <algorithm>
#include <functional>
#include <iostream>
#include <vector>
#include <omp.h>

#pragma omp declare reduction(+ : std::vector<double> : std::transform(omp_in.begin(),omp_in.end(),omp_out.begin(),omp_out.begin(),std::plus<double>())) initializer (omp_priv=omp_orig)


int check1(int size){
  std::vector<double> result(size,0.0);

  int npart=100;
  int n;
  {
    int i;
// Works in GCC 6.3, Intel 17.0:
#pragma omp parallel for private(i,n) reduction(+:result)
    for (n=0; n<npart; n++){
      for (i=0; i<size; i++){
	result  += i;
      }
    }
  }
  int fail=0;
  for (int i=0; i<size; i++) {
    if (result != i*100){
      fail=1;
    } 
    std::cout <<"i="<< i<<" "<<result<<std::endl;
  }
  return fail;
}

int check2(int size){
  std::vector<double> result(size,0.0);
  std::vector<double> result2(size,0.0);

  int npart=100;
  int n;
  {
    int i;
// Works in GCC 6.3, fails in Intel 17.0:
//#pragma omp parallel for private(i,n) reduction(+:result),reduction(+:result2)
// Works in GCC 6.3, fails in Intel 17.0:
#pragma omp parallel for private(i,n) reduction(+:result,result2)
    for (n=0; n<npart; n++){
      for (i=0; i<size; i++){
	result  += i;
	result2 += i;
      }
    }
  }
  int fail=0;
  for (int i=0; i<size; i++) {
    if (result != i*100 || result2 != i*100){
      fail=1;
    }
    std::cout <<"i="<< i<<" "<<result<<" "<<result2<<std::endl;
  }
  return fail;
}



int main(int argc, char *argv[]) {

  int size;

  if (argc < 2)
    size = 10;
  else
    size = atoi(argv[1]);

  int fail1=check1(size);
  int fail2=check2(size);

  return fail1+fail2;
}

I guess the question would

TimP — Tue, 02 May 2017 19:00:53 GMT

I guess the question would get more rapid attention on the C++ forum or with a submission to supporttickets.com

I'll try this myself when I get a chance.

Sean_D_ — Tue, 02 May 2017 19:29:08 GMT

Thank you I just resubmitted to the C++ forum.

topic multiple reduction on stl::vector gets wrong results in Intel® Moderncode for Parallel Architectures

multiple reduction on stl::vector gets wrong results

I guess the question would