Edit

Share via

Compiler Warning (level 4) C4938

  • Article

'var' : Floating point reduction variable may cause inconsistent results under /fp:strict or #pragma fenv_access

You should not use /fp:strict or fenv_access with OpenMP floating-point reductions, because the sum is computed in a different order. Thus, results can differ from the results without /openmp.

The following sample generates C4938:

// C4938.cpp
// compile with: /openmp /W4 /fp:strict /c
// #pragma fenv_access(on)
extern double *a;

double test(int first, int last) {
   double sum = 0.0;
   #pragma omp parallel for reduction(+: sum)   // C4938
   for (int i = first ; i <= last ; ++i)
      sum += a[i];
   return sum;
}

Without explicit parallelization, the sum is computed as follows:

sum = a[first] + a[first + 1] + ... + a[last];

With explicit parallelization (and two threads), the sum is computed as follows:

sum1 = a[first] + ... a[first + last / 2];
sum2 = a[(first + last / 2) + 1] + ... a[last];
sum = sum1 + sum2;