OpenMP parallel for loop and std::vector reduction

Asked May 07 '18 at 17:49

Active Sep 07 '18 at 22:41

Viewed 924 times

I have the following simplified code fragment that produces slightly different results at each execution. Results also differ from the serial version. Could anyone please tell me what is wrong with it ?

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

vector<float> vec0;
vec0.resize(M);
std::fill (vec0.begin(),vec0.end(),0);

#pragma omp parallel for reduction(vec_float_plus : vec0) 

for (int i0=0;i0<M;i0++)
{
    for (int j0=0;j0<M;j0++)
    {
        vec0[j0]+=A[i0]*B[i0]*C[i0]; // A, B, C are vectors of size M
    } 
}

edited May 08 '18 at 14:57

asked May 07 '18 at 17:49

user1364012

1

Maybe it is because [floating point addition is not associative](https://stackoverflow.com/a/10371890/9385966). Not enough details to tell, we would need an [MCVE](https://stackoverflow.com/help/mcve). – gflegar May 07 '18 at 17:56
1

Quantify *slightly different results*. – High Performance Mark May 07 '18 at 18:32
@High Performance Mark difference is around 0.1 % – user1364012 May 07 '18 at 19:50
@Goran Flegar I will make it as soon as possible – user1364012 May 07 '18 at 19:52
What compiler are you using? https://stackoverflow.com/q/50222760/2542702 – Z boson May 08 '18 at 07:35
@Z boson g++ 7.3.0 – user1364012 May 08 '18 at 12:10
Adding float operations are very prone to underflow approximations. Done the wrong way you can easily show that sum 1/n is a finite sum when we all now that it converges to infinity… maybe your data is very heterogeneous (1,0 to 1,0e-15) and the way you add data is important hence the difference between serial and parallel – PilouPili Sep 08 '18 at 18:34

OpenMP parallel for loop and std::vector reduction

0 Answers0