Kokkos.cpp

Kokkos::TeamPolicy<Kokkos::Cuda> execution_policy(total_amount_of_work, 
                                  outer_parallelism_batch_size,
                                  inner_parallelism_batch_size);

Kokkos::parallel_for("implement_psychics", execution_policy, 
                     KOKKOS_LAMBDA (team_member outer_handle) {
  // some outer level work
  double sum = psychics;                     
  parallel_reduce (ThreadVectorRange (outer_handle, loop_count),
          [=] (int& i, Scalar& lsum) {
            // some inner level work
            lsum += ...;
          }, sum);
                       
});