On 2017-11-21 16:50 +0800, Xi Ruoyao wrote:
> On 2017-11-21 15:49 +0800, Nan Xiao wrote:
> >
> > #include <omp.h>
> > #include <stdio.h>
> >
> > int main(void) {
> > #pragma omp parallel for
> > for (auto i = 0; i < 10; i++) {
> > int sum = 0;
> > #pragma omp taskloop shared(sum)
> > for (auto j = 0; j < 1000000; j++) {
> > sum += j;
> > }
> > printf("%d\n", sum);
> > }
> > return 0;
> > }
>
> There are two bugs in your code. First, signed overflow is an undefined
> behaviour and may generate arbitary result. Second, the access to shared
> variable sum is racing, the result may vary with scheduling.
Fix:
#pragma omp parallel for
for (auto i = 0; i < 10; i++) {
long long sum = 0;
#pragma omp taskloop shared(sum)
for (auto j = 0; j < 1000000; j++) {
__atomic_add_fetch(&sum, j, __ATOMIC_RELAXED);
}
printf("%lld\n", sum);
}
This would generate "lock addq" instruction for "sum", instead of loading
it into a register.
--
Xi Ruoyao <ryxi@xxxxxxxxxxxxxxxxx>
School of Aerospace Science and Technology, Xidian University
