Rework CUDA cache config using carveout calculations

[crtrott]

This will generally rely on the CUDA runtime to figure out best cache configuration, without us setting it manually anymore. We only set it manually if and when someone requests restricted occupancy. However, I did test the carveout scheme when always manually setting, and it did basically replicate the performance obtained by not setting anything for LAMMPS - and beats what we had before which was more limited. I also wrote a test code to check that the occupancy limitation works more or less as expected. Note that this is fairly iffy since there are many moving parts.

#include <Kokkos_Core.hpp>
#include <cmath>

void check_range(int N, int occupancy, int work, int size) {
  Kokkos::View<int> count("C");
  Kokkos::View<double**> workarray("A",N,size);

  auto lambda = KOKKOS_LAMBDA(int i, int& lmax) {
    int id = Kokkos::atomic_fetch_add(&count(), 1);
    for(int j=1; j<work; j++) workarray(i,j%size) += workarray(i, (j-1)%size) + j;
    Kokkos::atomic_add(&count(), -1);
    if(id > lmax) lmax = id;
  };

  Kokkos::RangePolicy<> p(0,N);
  auto occ_p = Kokkos::Experimental::prefer(p, Kokkos::Experimental::DesiredOccupancy{occupancy});
  int max_val;
  Kokkos::parallel_reduce(occ_p, lambda, Kokkos::Max<int>(max_val));
  printf("Occ: %i Max: %i\n",occupancy, max_val);
}

void check_team(int N, int occupancy, int work, int size, int team_size, int shmem) {
  Kokkos::View<int> count("C");
  Kokkos::View<double**> workarray("A",N,size);

  auto lambda = KOKKOS_LAMBDA(const typename Kokkos::TeamPolicy<>::member_type& team, int& lmax) {
    int i = team.league_rank()*team_size+team.team_rank();
    int id = Kokkos::atomic_fetch_add(&count(), 1);
    for(int j=1; j<work; j++) workarray(i,j%size) += workarray(i, (j-1)%size) + j;
    Kokkos::atomic_add(&count(), -1);
    if(id > lmax) lmax = id;
  };

  Kokkos::TeamPolicy<> p(N/team_size, team_size);
  p.set_scratch_size(0, Kokkos::PerTeam(shmem));
  auto occ_p = Kokkos::Experimental::prefer(p, Kokkos::Experimental::DesiredOccupancy{occupancy});
  int max_val;
  Kokkos::parallel_reduce(occ_p, lambda, Kokkos::Max<int>(max_val));
  printf("Occ: %i Max: %i\n",occupancy, max_val);
}
int main(int argc, char* argv[]) {
  Kokkos::initialize(argc, argv);
  {
    int N = argc > 1 ? atoi(argv[1]) : 1000000;
    int work = argc > 2 ? atoi(argv[2]) : 10000;
    int size = argc > 3 ? atoi(argv[3]) : 100;
    int occupancy = argc > 4 ? atoi(argv[4]) : 100;
    int team_size = argc > 5 ? atoi(argv[5]) : 128;
    int shmem = argc > 6 ? atoi(argv[6]) : 0;
    check_range(N,occupancy,work,size);
    check_team(N, occupancy, work, size, team_size, shmem);
  }
  Kokkos::finalize();
}

We may consider adding this test, however it requires that the GPU is used exclusively.

[PhilMiller]

Tagged as blocking in lieu of #4295

[dalg24]

Make sure you add a description

[crtrott]

Retest this please.

[dalg24]

I wish you did not interleave out and in parameters...

[dalg24]

Avoid "magic" numbers. Give it a name.

[crtrott]

null_point_eight?