Addition with std::complex constructor in openacc parallel loop produces incorrect array value

Accelerated Computing HPC Compilers nvc, nvc++ and nvfortran
1

There’s a file openacc_test.cpp:

#include <complex>
#include <iostream>
#include <vector>

using t_complex = std::complex<double>;

int main()
{
    auto n = std::size_t(100);
    auto v = std::vector<t_complex>(n);
    auto data = v.data();
#pragma acc data copy(data [0:n])
    {
#pragma acc kernels
        for (auto k = std::size_t(0); k < n; ++k)
        {
            data[k] = t_complex(0.0, 0.0);
            data[k] += t_complex(double(k), double(k));
        }
    }
    std::cout << data[2] << "\n";
}

Why does it print (96,96) instead of (2,2) when compiled with -acc?

$ nvc++ --version

nvc++ 21.2-0 LLVM 64-bit target on x86-64 Linux -tp zen 
NVIDIA Compilers and Tools
Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
$ g++ openacc_test.cpp && ./a.out 
(2,2)
$ nvc++ -acc openacc_test.cpp  -Minfo=accel && ./a.out
main:
     13, Generating copy(data[:n]) [if not already present]
         Generating implicit copy(_T20074136_7370) [if not already present]
         Generating implicit copyin(_T19259568_7370) [if not already present]
     15, Loop is parallelizable
         Generating Tesla code
         15, #pragma acc loop gang, vector(128) /* blockIdx.x threadIdx.x */
std::complex<double>::complex(double, double):
      1, include "complex"
          38, include "complex"
             1234, Generating implicit acc routine seq
                   Generating acc routine seq
                   Generating Tesla code
std::complex<double>::complex(double, double) [subobject]:
      0, Generating implicit acc routine seq
         Generating acc routine seq
         Generating Tesla code
std::complex<double>::__rep() const:
      1, include "complex"
          38, include "complex"
             1361, Generating implicit acc routine seq
                   Generating acc routine seq
                   Generating Tesla code
std::complex<double>& std::complex<double>::operator +=<double>(const std::complex<T1>&):
      1, include "complex"
          38, include "complex"
             1330, Generating implicit acc routine seq
                   Generating acc routine seq
                   Generating Tesla code
(96,96)
2

Looks to me like we’re hoisting the declaration of “t_complex(double(k), double(k))” and then copying in as a shared variable thus causing a race condition.

I’ve add a problem report, TPR #29909, and sent it to engineering for further investigation.

Thanks for the report!
Mat