parallel_reduce with RDC crashes linker

[ibaned]

Just like the situation described in this commit message: a98c130, I am unable to link any code that uses parallel_reduce with Kokkos and CUDA, if and only if relocatable device code (RDC) is being used. The linker segfaults:

nvcc error   : 'nvlink' died due to signal 11 (Invalid memory reference)

Several standalone example programs and one large library confirm that parallel_reduce+RDC is the problematic combination. Also replicated with many Kokkos commits going back to 2015, unable to find a commit without this behavior.

If this is supposed to work, here is more information to reproduce on Shannon:

../generate_makefile.bash \
--prefix=$HOME/kokkos-install \
--with-cuda=/home/projects/x86-64-sandybridge-nvidia/cuda/7.5.7 \
--arch=Kepler37 \
--with-cuda-options="enable_lambda,rdc" \
--compiler=$HOME/kokkos/config/nvcc_wrapper

#include <Kokkos_Core.hpp>

int main(int argc, char** argv)
{
  Kokkos::initialize(argc, argv);
  int n = 5;
  double sum = 0;
  Kokkos::parallel_reduce(n, KOKKOS_LAMBDA(int i, double& update) {
      update += 42; }, sum);
  Kokkos::finalize();
}

[hcedwar]

This is a known bug in nvcc linker for all current versions of CUDA. A bug report has been filed with NVIDIA and a resolution should soon be available.

[ibaned]

I see. Until then, I suppose my best option is to try not to rely on RDC. parallel_reduce is essential to my algorithms, so that has to stay. Thank you.

[hcedwar]

This nvcc linker bug is fixed in CUDA 8; which will be publicly available this summer as announced at GPU-Tech conference on April 5, 2016.