src/gromacs/mdlib/domdec_network.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2008,2009,2010,2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35
  36 #include "config.h"
  37
  38 #include <string.h>
  39 #include "types/commrec.h"
  40 #include "domdec_network.h"
  41
  42 #include "gromacs/utility/gmxmpi.h"
  43
  44
  45 #define DDMASTERRANK(dd)   (dd->masterrank)
  46
  47
  48 void dd_sendrecv_int(const gmx_domdec_t gmx_unused *dd,
  49                      int gmx_unused ddimind, int gmx_unused direction,
  50                      int gmx_unused *buf_s, int gmx_unused n_s,
  51                      int gmx_unused *buf_r, int gmx_unused n_r)
  52 {
  53 #ifdef GMX_MPI
  54     int        rank_s, rank_r;
  55     MPI_Status stat;
  56
  57     rank_s = dd->neighbor[ddimind][direction == dddirForward ? 0 : 1];
  58     rank_r = dd->neighbor[ddimind][direction == dddirForward ? 1 : 0];
  59
  60     if (n_s && n_r)
  61     {
  62         MPI_Sendrecv(buf_s, n_s*sizeof(int), MPI_BYTE, rank_s, 0,
  63                      buf_r, n_r*sizeof(int), MPI_BYTE, rank_r, 0,
  64                      dd->mpi_comm_all, &stat);
  65     }
  66     else if (n_s)
  67     {
  68         MPI_Send(    buf_s, n_s*sizeof(int), MPI_BYTE, rank_s, 0,
  69                      dd->mpi_comm_all);
  70     }
  71     else if (n_r)
  72     {
  73         MPI_Recv(    buf_r, n_r*sizeof(int), MPI_BYTE, rank_r, 0,
  74                      dd->mpi_comm_all, &stat);
  75     }
  76
  77 #endif
  78 }
  79
  80 void dd_sendrecv_real(const gmx_domdec_t gmx_unused *dd,
  81                       int gmx_unused ddimind, int gmx_unused direction,
  82                       real gmx_unused *buf_s, int gmx_unused n_s,
  83                       real gmx_unused *buf_r, int gmx_unused n_r)
  84 {
  85 #ifdef GMX_MPI
  86     int        rank_s, rank_r;
  87     MPI_Status stat;
  88
  89     rank_s = dd->neighbor[ddimind][direction == dddirForward ? 0 : 1];
  90     rank_r = dd->neighbor[ddimind][direction == dddirForward ? 1 : 0];
  91
  92     if (n_s && n_r)
  93     {
  94         MPI_Sendrecv(buf_s, n_s*sizeof(real), MPI_BYTE, rank_s, 0,
  95                      buf_r, n_r*sizeof(real), MPI_BYTE, rank_r, 0,
  96                      dd->mpi_comm_all, &stat);
  97     }
  98     else if (n_s)
  99     {
 100         MPI_Send(    buf_s, n_s*sizeof(real), MPI_BYTE, rank_s, 0,
 101                      dd->mpi_comm_all);
 102     }
 103     else if (n_r)
 104     {
 105         MPI_Recv(    buf_r, n_r*sizeof(real), MPI_BYTE, rank_r, 0,
 106                      dd->mpi_comm_all, &stat);
 107     }
 108
 109 #endif
 110 }
 111
 112 void dd_sendrecv_rvec(const gmx_domdec_t gmx_unused *dd,
 113                       int gmx_unused ddimind, int gmx_unused direction,
 114                       rvec gmx_unused *buf_s, int gmx_unused n_s,
 115                       rvec gmx_unused *buf_r, int gmx_unused n_r)
 116 {
 117 #ifdef GMX_MPI
 118     int        rank_s, rank_r;
 119     MPI_Status stat;
 120
 121     rank_s = dd->neighbor[ddimind][direction == dddirForward ? 0 : 1];
 122     rank_r = dd->neighbor[ddimind][direction == dddirForward ? 1 : 0];
 123
 124     if (n_s && n_r)
 125     {
 126         MPI_Sendrecv(buf_s[0], n_s*sizeof(rvec), MPI_BYTE, rank_s, 0,
 127                      buf_r[0], n_r*sizeof(rvec), MPI_BYTE, rank_r, 0,
 128                      dd->mpi_comm_all, &stat);
 129     }
 130     else if (n_s)
 131     {
 132         MPI_Send(    buf_s[0], n_s*sizeof(rvec), MPI_BYTE, rank_s, 0,
 133                      dd->mpi_comm_all);
 134     }
 135     else if (n_r)
 136     {
 137         MPI_Recv(    buf_r[0], n_r*sizeof(rvec), MPI_BYTE, rank_r, 0,
 138                      dd->mpi_comm_all, &stat);
 139     }
 140
 141 #endif
 142 }
 143
 144 void dd_sendrecv2_rvec(const gmx_domdec_t gmx_unused *dd,
 145                        int gmx_unused ddimind,
 146                        rvec gmx_unused *buf_s_fw, int gmx_unused n_s_fw,
 147                        rvec gmx_unused *buf_r_fw, int gmx_unused n_r_fw,
 148                        rvec gmx_unused *buf_s_bw, int gmx_unused n_s_bw,
 149                        rvec gmx_unused *buf_r_bw, int gmx_unused n_r_bw)
 150 {
 151 #ifdef GMX_MPI
 152     int         rank_fw, rank_bw, nreq;
 153     MPI_Request req[4];
 154     MPI_Status  stat[4];
 155
 156     rank_fw = dd->neighbor[ddimind][0];
 157     rank_bw = dd->neighbor[ddimind][1];
 158
 159     if (!dd->bSendRecv2)
 160     {
 161         /* Try to send and receive in two directions simultaneously.
 162          * Should be faster, especially on machines
 163          * with full 3D communication networks.
 164          * However, it could be that communication libraries are
 165          * optimized for MPI_Sendrecv and non-blocking MPI calls
 166          * are slower.
 167          * SendRecv2 can be turned on with the env.var. GMX_DD_SENDRECV2
 168          */
 169         nreq = 0;
 170         if (n_r_fw)
 171         {
 172             MPI_Irecv(buf_r_fw[0], n_r_fw*sizeof(rvec), MPI_BYTE,
 173                       rank_bw, 0, dd->mpi_comm_all, &req[nreq++]);
 174         }
 175         if (n_r_bw)
 176         {
 177             MPI_Irecv(buf_r_bw[0], n_r_bw*sizeof(rvec), MPI_BYTE,
 178                       rank_fw, 1, dd->mpi_comm_all, &req[nreq++]);
 179         }
 180         if (n_s_fw)
 181         {
 182             MPI_Isend(buf_s_fw[0], n_s_fw*sizeof(rvec), MPI_BYTE,
 183                       rank_fw, 0, dd->mpi_comm_all, &req[nreq++]);
 184         }
 185         if (n_s_bw)
 186         {
 187             MPI_Isend(buf_s_bw[0], n_s_bw*sizeof(rvec), MPI_BYTE,
 188                       rank_bw, 1, dd->mpi_comm_all, &req[nreq++]);
 189         }
 190         if (nreq)
 191         {
 192             MPI_Waitall(nreq, req, stat);
 193         }
 194     }
 195     else
 196     {
 197         /* Communicate in two ordered phases.
 198          * This is slower, even on a dual-core Opteron cluster
 199          * with a single full-duplex network connection per machine.
 200          */
 201         /* Forward */
 202         MPI_Sendrecv(buf_s_fw[0], n_s_fw*sizeof(rvec), MPI_BYTE, rank_fw, 0,
 203                      buf_r_fw[0], n_r_fw*sizeof(rvec), MPI_BYTE, rank_bw, 0,
 204                      dd->mpi_comm_all, &stat[0]);
 205         /* Backward */
 206         MPI_Sendrecv(buf_s_bw[0], n_s_bw*sizeof(rvec), MPI_BYTE, rank_bw, 0,
 207                      buf_r_bw[0], n_r_bw*sizeof(rvec), MPI_BYTE, rank_fw, 0,
 208                      dd->mpi_comm_all, &stat[0]);
 209     }
 210 #endif
 211 }
 212
 213 /* IBM's BlueGene(/L) MPI_Bcast dereferences the data pointer
 214  * even when 0 == nbytes, so we protect calls to it on BlueGene.
 215  * Fortunately dd_bcast() and dd_bcastc() are only
 216  * called during DD setup and partition.
 217  */
 218
 219 void dd_bcast(gmx_domdec_t gmx_unused *dd, int gmx_unused nbytes, void gmx_unused *data)
 220 {
 221 #ifdef GMX_MPI
 222 #ifdef GMX_BLUEGENE
 223     if (nbytes > 0)
 224     {
 225 #endif
 226     MPI_Bcast(data, nbytes, MPI_BYTE,
 227               DDMASTERRANK(dd), dd->mpi_comm_all);
 228 #ifdef GMX_BLUEGENE
 229 }
 230 #endif
 231 #endif
 232 }
 233
 234 void dd_bcastc(gmx_domdec_t *dd, int nbytes, void *src, void *dest)
 235 {
 236     if (DDMASTER(dd))
 237     {
 238         memcpy(dest, src, nbytes);
 239     }
 240 #ifdef GMX_MPI
 241 #ifdef GMX_BLUEGENE
 242     if (nbytes > 0)
 243     {
 244 #endif
 245     MPI_Bcast(dest, nbytes, MPI_BYTE,
 246               DDMASTERRANK(dd), dd->mpi_comm_all);
 247 #ifdef GMX_BLUEGENE
 248 }
 249 #endif
 250 #endif
 251 }
 252
 253 void dd_scatter(gmx_domdec_t gmx_unused *dd, int gmx_unused nbytes, void gmx_unused *src, void gmx_unused *dest)
 254 {
 255 #ifdef GMX_MPI
 256     MPI_Scatter(src, nbytes, MPI_BYTE,
 257                 dest, nbytes, MPI_BYTE,
 258                 DDMASTERRANK(dd), dd->mpi_comm_all);
 259 #endif
 260 }
 261
 262 void dd_gather(gmx_domdec_t gmx_unused *dd, int gmx_unused nbytes, void gmx_unused *src, void gmx_unused *dest)
 263 {
 264 #ifdef GMX_MPI
 265     MPI_Gather(src, nbytes, MPI_BYTE,
 266                dest, nbytes, MPI_BYTE,
 267                DDMASTERRANK(dd), dd->mpi_comm_all);
 268 #endif
 269 }
 270
 271 void dd_scatterv(gmx_domdec_t gmx_unused *dd,
 272                  int gmx_unused *scounts, int gmx_unused *disps, void gmx_unused *sbuf,
 273                  int gmx_unused rcount, void gmx_unused *rbuf)
 274 {
 275 #ifdef GMX_MPI
 276     int dum;
 277
 278     if (rcount == 0)
 279     {
 280         /* MPI does not allow NULL pointers */
 281         rbuf = &dum;
 282     }
 283     MPI_Scatterv(sbuf, scounts, disps, MPI_BYTE,
 284                  rbuf, rcount, MPI_BYTE,
 285                  DDMASTERRANK(dd), dd->mpi_comm_all);
 286 #endif
 287 }
 288
 289 void dd_gatherv(gmx_domdec_t gmx_unused *dd,
 290                 int gmx_unused scount, void gmx_unused *sbuf,
 291                 int gmx_unused *rcounts, int gmx_unused *disps, void gmx_unused *rbuf)
 292 {
 293 #ifdef GMX_MPI
 294     int dum;
 295
 296     if (scount == 0)
 297     {
 298         /* MPI does not allow NULL pointers */
 299         sbuf = &dum;
 300     }
 301     MPI_Gatherv(sbuf, scount, MPI_BYTE,
 302                 rbuf, rcounts, disps, MPI_BYTE,
 303                 DDMASTERRANK(dd), dd->mpi_comm_all);
 304 #endif
 305 }