Wrong result with openMP reduction pragma

Question

I have this code that was working for years (and is still working when using some random compilers).

What we expect is to have the same result in sequential and in parallel execution.

The symptom is that at each execution, the parallel execution produces another result.

#include <omp.h>
#include <stdio.h>
#include <stdlib.h>

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

    int i, N, j, sum;
    int ** A;

    sum=0;
    N=1000;

    A=(int**)malloc(N*sizeof(int*));
    for (i=0;i<N;i++) {
        A[i] = (int*)malloc(N *sizeof(int));
    }

    for (i=0; i<N; ++i) {
        for (j=0; j<N; ++j) {
            A[i][j]=i+j;
            sum+=A[i][j];
        }
    }

    printf("Total sum = %d \n",sum);
    sum=0;


    #pragma omp parallel for reduction(+:sum)
    for (i=0; i<N; ++i) {
        for (j=0; j<N; ++j) {
            sum += A[i][j];
        }
    }

    printf("Total sum = %d \n",sum);

    for (i=0;i<N;i++){ free(A[i]);}
    free(A);

    return 0;
}

We compile it like that:

gcc -fopenmp reduction.c

And run it like that:

./a.out
Total sum = 999000000
Total sum = 822136991

It's working with icc.

Edit: if we use optimization -O3 with Gcc it's working also.

Use this `#pragma omp parallel for reduction(+:sum) private(j)`. — Z boson, Sep 12 '14 at 10:40
@Zboson thanks it seems correct. If you create an answer I'll accept it. — Yann Sagon, Sep 14 '14 at 19:04

score 1 · Answer 1 · edited May 23 '17 at 12:28

1

The problem is that you have a nested loop, and the pragma is only applying for the outer one. You need to use a collapse clause. You can read about it this question and in this site. The program works correctly if you replace your #pragma line with:

#pragma omp parallel for reduction(+:sum) collapse(2)

edited May 23 '17 at 12:28

Community

1
1

answered Sep 12 '14 at 10:23

jdehesa

58,456
7
77
121

2

Or he could just make `j` private. – Z boson Sep 12 '14 at 10:40

score 1 · Accepted Answer · answered Sep 14 '14 at 19:54

Here are three ways to fix your code

Explicitly make j private

#pragma omp parallel for reduction(+:sum) private(j)
for (i=0; i<N; ++i) {
    for (j=0; j<N; ++j) {
        sum += A[i][j];
    }
}

Change your code and define i and j inside a parallel region

#pragma omp parallel reduction(+:sum)
{
    int i,j;
    #pragma omp for
    for (i=0; i<N; ++i) {
        for (j=0; j<N; ++j) {
            sum += A[i][j];
        }
    }
}

Use C99 (or GNU99) and change your code to

#pragma omp parallel for reduction(+:sum)
for (int i=0; i<N; ++i) {
    for (int j=0; j<N; ++j) {
        sum += A[i][j];
    }
}

Wrong result with openMP reduction pragma

2 Answers2