MVAPICH2 multi-gpu initialization

mvapich2-cuda-init.c

void device_binding ()
{

  int local_rank, num_local_procs;
  int dev_count, use_dev_count, my_dev_id;
  char *str;

  if ((str = getenv ("MV2_COMM_WORLD_LOCAL_RANK")) != NULL)
    {
      local_rank = atoi (str);
      printf ("MV2_COMM_WORLD_LOCAL_RANK %s\n", str);
    }

  if ((str = getenv ("MPISPAWN_LOCAL_NPROCS")) != NULL)
    {
      num_local_procs = atoi (str);
      printf ("MPISPAWN_LOCAL_NPROCS %s\n", str);
    }

  cudaGetDeviceCount (&dev_count);
  if ((str = getenv ("NUM_GPU_DEVICES")) != NULL)
    {
      use_dev_count = atoi (str);
      printf ("NUM_GPU_DEVICES %s\n", str);
    }
  else
    {
      use_dev_count = dev_count;
    }

  my_dev_id = local_rank % use_dev_count;
  printf ("local rank = %d dev id = %d\n", local_rank, my_dev_id);
  cudaSetDevice (my_dev_id);
}