C++ AtomVec::unpack_exchange方法代码示例

//.........这里部分代码省略.........
  //   new ghosts are created in borders()
  // map_set() is done at end of borders()
  // clear ghost count and any ghost bonus data internal to AtomVec

  if (map_style) atom->map_clear();
  atom->nghost = 0;
  atom->avec->clear_bonus();

  // insure send buf is large enough for single atom
  // bufextra = max size of one atom = allowed overflow of sendbuf
  // fixes can change per-atom size requirement on-the-fly

  int bufextra_old = bufextra;
  maxexchange = maxexchange_atom + maxexchange_fix;
  bufextra = maxexchange + BUFEXTRA;
  if (bufextra > bufextra_old)
    memory->grow(buf_send,maxsend+bufextra,"comm:buf_send");

  // subbox bounds for orthogonal or triclinic

  if (triclinic == 0) {
    sublo = domain->sublo;
    subhi = domain->subhi;
  } else {
    sublo = domain->sublo_lamda;
    subhi = domain->subhi_lamda;
  }

  // loop over dimensions

  int dimension = domain->dimension;

  for (int dim = 0; dim < dimension; dim++) {

    // fill buffer with atoms leaving my box, using < and >=
    // when atom is deleted, fill it in with last atom

    x = atom->x;
    lo = sublo[dim];
    hi = subhi[dim];
    nlocal = atom->nlocal;
    i = nsend = 0;

    while (i < nlocal) {
      if (x[i][dim] < lo || x[i][dim] >= hi) {
        if (nsend > maxsend) grow_send(nsend,1);
        nsend += avec->pack_exchange(i,&buf_send[nsend]);
        avec->copy(nlocal-1,i,1);
        nlocal--;
      } else i++;
    }
    atom->nlocal = nlocal;

    // send/recv atoms in both directions
    // send size of message first so receiver can realloc buf_recv if needed
    // if 1 proc in dimension, no send/recv
    //   set nrecv = 0 so buf_send atoms will be lost
    // if 2 procs in dimension, single send/recv
    // if more than 2 procs in dimension, send/recv to both neighbors

    if (procgrid[dim] == 1) nrecv = 0;
    else {
      MPI_Sendrecv(&nsend,1,MPI_INT,procneigh[dim][0],0,
                   &nrecv1,1,MPI_INT,procneigh[dim][1],0,world,MPI_STATUS_IGNORE);
      nrecv = nrecv1;
      if (procgrid[dim] > 2) {
        MPI_Sendrecv(&nsend,1,MPI_INT,procneigh[dim][1],0,
                     &nrecv2,1,MPI_INT,procneigh[dim][0],0,world,MPI_STATUS_IGNORE);
        nrecv += nrecv2;
      }
      if (nrecv > maxrecv) grow_recv(nrecv);

      MPI_Irecv(buf_recv,nrecv1,MPI_DOUBLE,procneigh[dim][1],0,
                world,&request);
      MPI_Send(buf_send,nsend,MPI_DOUBLE,procneigh[dim][0],0,world);
      MPI_Wait(&request,MPI_STATUS_IGNORE);

      if (procgrid[dim] > 2) {
        MPI_Irecv(&buf_recv[nrecv1],nrecv2,MPI_DOUBLE,procneigh[dim][0],0,
                  world,&request);
        MPI_Send(buf_send,nsend,MPI_DOUBLE,procneigh[dim][1],0,world);
        MPI_Wait(&request,MPI_STATUS_IGNORE);
      }
    }

    // check incoming atoms to see if they are in my box
    // if so, add to my list
    // box check is only for this dimension,
    //   atom may be passed to another proc in later dims

    m = 0;
    while (m < nrecv) {
      value = buf_recv[m+dim+1];
      if (value >= lo && value < hi) m += avec->unpack_exchange(&buf_recv[m]);
      else m += static_cast<int> (buf_recv[m]);
    }
  }

  if (atom->firstgroupname) atom->first_reorder();
}

void CommCuda::exchange_cuda()
{
  int i,m,nsend,nrecv,nrecv1,nrecv2,nlocal;
  double lo,hi,value;
  double **x;
  double *sublo,*subhi,*buf;
  MPI_Request request;
  MPI_Status status;
  AtomVec *avec = atom->avec;
    timespec time1,time2,time3;

  // clear global->local map for owned and ghost atoms
  // b/c atoms migrate to new procs in exchange() and 
  // new ghosts are created in borders()
  // map_set() is done at end of borders()

  
  if(map_style) cuda->cu_tag->download();

  if (map_style) atom->map_clear();

  // subbox bounds for orthogonal or triclinic

  if (triclinic == 0) {
    sublo = domain->sublo;
    subhi = domain->subhi;
  } else {
    sublo = domain->sublo_lamda;
    subhi = domain->subhi_lamda;
  }

  // loop over dimensions
	
  for (int dim = 0; dim < 3; dim++) {
    // fill buffer with atoms leaving my box, using < and >=
    // when atom is deleted, fill it in with last atom

  	cuda->shared_data.exchange_dim=dim;
    
    nlocal = atom->nlocal;
    avec->maxsend=&maxsend;
    nsend=avec->pack_exchange(dim,(double*) &buf_send);
    nlocal = atom->nlocal;


    atom->nlocal = nlocal;

    // send/recv atoms in both directions
    // if 1 proc in dimension, no send/recv, set recv buf to send buf
    // if 2 procs in dimension, single send/recv
    // if more than 2 procs in dimension, send/recv to both neighbors

 clock_gettime(CLOCK_REALTIME,&time1);
 
    if (procgrid[dim] == 1) {
      nrecv = nsend;
      buf = buf_send;

    } else {
      MPI_Sendrecv(&nsend,1,MPI_INT,procneigh[dim][0],0,
		   &nrecv1,1,MPI_INT,procneigh[dim][1],0,world,&status);
      nrecv = nrecv1;
      if (procgrid[dim] > 2) {
	MPI_Sendrecv(&nsend,1,MPI_INT,procneigh[dim][1],0,
		     &nrecv2,1,MPI_INT,procneigh[dim][0],0,world,&status);
	nrecv += nrecv2;
      }
      if (nrecv+1 > maxrecv) grow_recv(nrecv+1);
      
      MPI_Irecv(buf_recv,nrecv1,MPI_DOUBLE,procneigh[dim][1],0,
		world,&request);
      MPI_Send(buf_send,nsend,MPI_DOUBLE,procneigh[dim][0],0,world);
      MPI_Wait(&request,&status);
      
      if (procgrid[dim] > 2) {
	MPI_Irecv(&buf_recv[nrecv1],nrecv2,MPI_DOUBLE,procneigh[dim][0],0,
		  world,&request);
	MPI_Send(buf_send,nsend,MPI_DOUBLE,procneigh[dim][1],0,world);
	MPI_Wait(&request,&status);

    	if((nrecv1==0)||(nrecv2==0)) buf_recv[nrecv]=0;
      }
      
      buf = buf_recv;
    }
	//printf("nsend: %i nrecv: %i\n",nsend,nrecv);
    // check incoming atoms to see if they are in my box
    // if so, add to my list
clock_gettime(CLOCK_REALTIME,&time2);
cuda->shared_data.cuda_timings.comm_exchange_mpi+=
      time2.tv_sec-time1.tv_sec+1.0*(time2.tv_nsec-time1.tv_nsec)/1000000000;

    if(nrecv)
    {
      avec->maxsend=&maxsend;
      avec->unpack_exchange(buf);
    }
  }

  if(atom->firstgroupname) cuda->downloadAll();
//.........这里部分代码省略.........