TeamPolicy doesn't throw then created with non-viable vector length and also doesn't backscale to viable one

[crtrott]

This code doesn't behave well for V>32 on NVIDIA GPUs:

#include<Kokkos_Core.hpp>

int main(int argc, char* argv[]) {
  Kokkos::initialize(argc,argv);
  {
     int T = (argc>1) ? atoi(argv[1]) : 16;
     int V = (argc>2) ? atoi(argv[2]) : 1;

     Kokkos::parallel_for(Kokkos::TeamPolicy<>(1,T,V), KOKKOS_LAMBDA (const Kokkos::TeamPolicy<>::member_type& team) {
       double val = 0;
       if(team.team_rank()==1) val = 5;
       team.team_reduce(Kokkos::Sum<double,Kokkos::DefaultExecutionSpace>(val));
       if(team.team_rank()==0) printf("%lf\n",val);
     });
  }
  Kokkos::finalize();
}

Probably the vector length should just be dropped down to the maximum possible, since this is more like a hint and the user doesn't have explicit access to vector lanes anyway.

[jeffmiles63]

Created PR #2067 to address this issue

[crtrott]

I think this should mostly not warn, and we treat the vector length as more something like a hint a la "might be a good vector length" what do you guys think?