// ---------------------------------------------------------------------------
// Allocate memory on host and device and copy constants to device
// ---------------------------------------------------------------------------
int eam_gpu_init(const int ntypes, double host_cutforcesq, 
                 int **host_type2rhor, int **host_type2z2r, int *host_type2frho,
                 double ***host_rhor_spline, double ***host_z2r_spline,
                 double ***host_frho_spline,
                 double rdr, double rdrho, int nrhor, 
                 int nrho, int nz2r, int nfrho, int nr, 
                 const int nlocal, const int nall, const int max_nbors, 
                 const int maxspecial, const double cell_size, 
                 int &gpu_mode, FILE *screen, int &fp_size) {
  EAMMF.clear();
  gpu_mode=EAMMF.device->gpu_mode();
  double gpu_split=EAMMF.device->particle_split();
  int first_gpu=EAMMF.device->first_device();
  int last_gpu=EAMMF.device->last_device();
  int world_me=EAMMF.device->world_me();
  int gpu_rank=EAMMF.device->gpu_rank();
  int procs_per_gpu=EAMMF.device->procs_per_gpu();

  // disable host/device split for now
  if (gpu_split != 1.0) 
    return -8;
    
  fp_size=sizeof(PRECISION);
    
  EAMMF.device->init_message(screen,"eam",first_gpu,last_gpu);

  bool message=false;
  if (EAMMF.device->replica_me()==0 && screen)
    message=true;

  if (message) {
    fprintf(screen,"Initializing GPU and compiling on process 0...");
    fflush(screen);
  }

  int init_ok=0;
  if (world_me==0)
    init_ok=EAMMF.init(ntypes, host_cutforcesq, host_type2rhor, host_type2z2r,
                       host_type2frho, host_rhor_spline, host_z2r_spline,
                       host_frho_spline, rdr, rdrho, nrhor, nrho, nz2r, nfrho,
                       nr, nlocal, nall, 300, maxspecial, cell_size, gpu_split,
                       screen);

  EAMMF.device->world_barrier();
  if (message)
    fprintf(screen,"Done.\n");

  for (int i=0; i<procs_per_gpu; i++) {
    if (message) {
      if (last_gpu-first_gpu==0)
        fprintf(screen,"Initializing GPU %d on core %d...",first_gpu,i);
      else
        fprintf(screen,"Initializing GPUs %d-%d on core %d...",first_gpu,
                last_gpu,i);
      fflush(screen);
    }
    if (gpu_rank==i && world_me!=0)
      init_ok=EAMMF.init(ntypes, host_cutforcesq, host_type2rhor, host_type2z2r,
                         host_type2frho, host_rhor_spline, host_z2r_spline,
                         host_frho_spline, rdr, rdrho, nrhor, nrho, nz2r, nfrho,
                         nr, nlocal, nall, 300, maxspecial, cell_size, 
                         gpu_split, screen);

    EAMMF.device->gpu_barrier();
    if (message) 
      fprintf(screen,"Done.\n");
  }
  if (message)
    fprintf(screen,"\n");

  if (init_ok==0)
    EAMMF.estimate_gpu_overhead();
  return init_ok;
}
void eam_fs_gpu_clear() {
  EAMFSMF.clear();
}
void eam_gpu_clear() {
  EAMMF.clear();
}
void eam_alloy_gpu_clear() {
  EAMALMF.clear();
}