https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121818
--- Comment #1 from Benjamin Schulz <schulz.benjamin at googlemail dot com> --- Created attachment 62321 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=62321&action=edit arraylibrary.tar.gz sourcecode of the arraylibrary. again, if you remove the reduction(+:tmp) in line 1610 of mathfunctions_mpi.h before the loop T tmp=0,temp4=0; #pragma omp target parallel for simd reduction(+:tmp) device(policy.devicenum) for (size_t k = 0; k < c; ++k) { const T tmp3=tL(c,k); tmp+= tmp3 * tmp3; } then the results come out correctly, but that should not be, that it computes correctly if i remove the reduciton, as this clearly is a a simple reduction. Originally, i had a teams distribute statement before, Then the behavior is the same. I removed that because I did not know whether it can reduce over the entire teams, but it does not matter apparently, the problem is there if i reduce even with parallel for simd, which supports the reduction statement in any case on target....
