print_gpu_binding Subroutine

   subroutine print_gpu_binding(requested_gpu)
      !! Per-rank one-shot diagnostic: world rank -> requested device num /
      !! actual device num reported by the OpenACC runtime.  Lets us
      !! confirm that mpirun is binding each rank to its own GPU rather
      !! than serialising N ranks on device 0.
      integer, intent(in) :: requested_gpu
      integer :: actual_dev

#ifdef RDB_HAS_OPENACC_RUNTIME
      ! Works on either backend: stdpar+OpenACC and -mp=gpu / -h omp
      ! both keep the OpenACC runtime live, and acc_get_device_num
      ! reports the same device that stdpar/-acc dispatches against.
      actual_dev = acc_get_device_num(acc_get_device_type())
#else
      actual_dev = -1
#endif
      write (output_unit, "(A,I0,A,I0,A,I0,A,I0,A,I0)") &
         "[gpu-bind] world_rank=", cached_rank, &
         "/", cached_size, &
         " compute_rank=", cached_compute_rank, &
         " requested_dev=", requested_gpu, &
         " actual_dev=", actual_dev
      if (cached_rank == 0) then
#ifdef RDB_CUDA_AWARE_MPI
         write (output_unit, "(A)") "[gpu-bind] halo path: GPU-direct (CUDA-aware MPI)"
#else
         write (output_unit, "(A)") "[gpu-bind] halo path: HOST-STAGED (rebuild with -DRDB_CUDA_AWARE_MPI=ON for GPU-direct)"
#endif
      end if
      flush (output_unit)
   end subroutine print_gpu_binding