2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
5 * Copyright (c) 2001-2004, The GROMACS development team,
6 * check out http://www.gromacs.org for more information.
7 * Copyright (c) 2012,2013, by the GROMACS development team, led by
8 * David van der Spoel, Berk Hess, Erik Lindahl, and including many
9 * others, as listed in the AUTHORS file in the top-level source
10 * directory and at http://www.gromacs.org.
12 * GROMACS is free software; you can redistribute it and/or
13 * modify it under the terms of the GNU Lesser General Public License
14 * as published by the Free Software Foundation; either version 2.1
15 * of the License, or (at your option) any later version.
17 * GROMACS is distributed in the hope that it will be useful,
18 * but WITHOUT ANY WARRANTY; without even the implied warranty of
19 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
20 * Lesser General Public License for more details.
22 * You should have received a copy of the GNU Lesser General Public
23 * License along with GROMACS; if not, see
24 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
25 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
27 * If you want to redistribute modifications to GROMACS, please
28 * consider that scientific software is very special. Version
29 * control is crucial - bugs must be traceable. We will be happy to
30 * consider code for inclusion in the official distribution, but
31 * derived work must not be called official GROMACS. Details are found
32 * in the README & COPYING files - if they are missing, get the
33 * official version at http://www.gromacs.org.
35 * To help us fund GROMACS development, we humbly ask that you cite
36 * the research papers on the package. Check out http://www.gromacs.org.
43 #include "gmx_fatal.h"
60 #include "mpelogging.h"
62 /* The source code in this file should be thread-safe.
63 Please keep it that way. */
65 gmx_bool gmx_mpi_initialized(void)
77 int gmx_setup(int *argc, char **argv, int *nnodes)
80 gmx_call("gmx_setup");
84 int resultlen; /* actual length of node name */
88 char mpi_hostname[MPI_MAX_PROCESSOR_NAME];
90 /* Call the MPI routines */
93 (void) fah_MPI_Init(argc, &argv);
95 (void) MPI_Init(argc, &argv);
98 (void) MPI_Comm_size( MPI_COMM_WORLD, &mpi_num_nodes );
99 (void) MPI_Comm_rank( MPI_COMM_WORLD, &mpi_my_rank );
100 (void) MPI_Get_processor_name( mpi_hostname, &resultlen );
104 /* MPE logging routines. Get event IDs from MPE: */
106 ev_timestep1 = MPE_Log_get_event_number( );
107 ev_timestep2 = MPE_Log_get_event_number( );
108 ev_force_start = MPE_Log_get_event_number( );
109 ev_force_finish = MPE_Log_get_event_number( );
110 ev_do_fnbf_start = MPE_Log_get_event_number( );
111 ev_do_fnbf_finish = MPE_Log_get_event_number( );
112 ev_ns_start = MPE_Log_get_event_number( );
113 ev_ns_finish = MPE_Log_get_event_number( );
114 ev_calc_bonds_start = MPE_Log_get_event_number( );
115 ev_calc_bonds_finish = MPE_Log_get_event_number( );
116 ev_global_stat_start = MPE_Log_get_event_number( );
117 ev_global_stat_finish = MPE_Log_get_event_number( );
118 ev_virial_start = MPE_Log_get_event_number( );
119 ev_virial_finish = MPE_Log_get_event_number( );
121 /* Shift related events */
122 ev_shift_start = MPE_Log_get_event_number( );
123 ev_shift_finish = MPE_Log_get_event_number( );
124 ev_unshift_start = MPE_Log_get_event_number( );
125 ev_unshift_finish = MPE_Log_get_event_number( );
126 ev_mk_mshift_start = MPE_Log_get_event_number( );
127 ev_mk_mshift_finish = MPE_Log_get_event_number( );
129 /* PME related events */
130 ev_pme_start = MPE_Log_get_event_number( );
131 ev_pme_finish = MPE_Log_get_event_number( );
132 ev_spread_on_grid_start = MPE_Log_get_event_number( );
133 ev_spread_on_grid_finish = MPE_Log_get_event_number( );
134 ev_sum_qgrid_start = MPE_Log_get_event_number( );
135 ev_sum_qgrid_finish = MPE_Log_get_event_number( );
136 ev_gmxfft3d_start = MPE_Log_get_event_number( );
137 ev_gmxfft3d_finish = MPE_Log_get_event_number( );
138 ev_solve_pme_start = MPE_Log_get_event_number( );
139 ev_solve_pme_finish = MPE_Log_get_event_number( );
140 ev_gather_f_bsplines_start = MPE_Log_get_event_number( );
141 ev_gather_f_bsplines_finish = MPE_Log_get_event_number( );
142 ev_reduce_start = MPE_Log_get_event_number( );
143 ev_reduce_finish = MPE_Log_get_event_number( );
144 ev_rscatter_start = MPE_Log_get_event_number( );
145 ev_rscatter_finish = MPE_Log_get_event_number( );
146 ev_alltoall_start = MPE_Log_get_event_number( );
147 ev_alltoall_finish = MPE_Log_get_event_number( );
148 ev_pmeredist_start = MPE_Log_get_event_number( );
149 ev_pmeredist_finish = MPE_Log_get_event_number( );
150 ev_init_pme_start = MPE_Log_get_event_number( );
151 ev_init_pme_finish = MPE_Log_get_event_number( );
152 ev_send_coordinates_start = MPE_Log_get_event_number( );
153 ev_send_coordinates_finish = MPE_Log_get_event_number( );
154 ev_update_fr_start = MPE_Log_get_event_number( );
155 ev_update_fr_finish = MPE_Log_get_event_number( );
156 ev_clear_rvecs_start = MPE_Log_get_event_number( );
157 ev_clear_rvecs_finish = MPE_Log_get_event_number( );
158 ev_update_start = MPE_Log_get_event_number( );
159 ev_update_finish = MPE_Log_get_event_number( );
160 ev_output_start = MPE_Log_get_event_number( );
161 ev_output_finish = MPE_Log_get_event_number( );
162 ev_sum_lrforces_start = MPE_Log_get_event_number( );
163 ev_sum_lrforces_finish = MPE_Log_get_event_number( );
164 ev_sort_start = MPE_Log_get_event_number( );
165 ev_sort_finish = MPE_Log_get_event_number( );
166 ev_sum_qgrid_start = MPE_Log_get_event_number( );
167 ev_sum_qgrid_finish = MPE_Log_get_event_number( );
169 /* Essential dynamics related events */
170 ev_edsam_start = MPE_Log_get_event_number( );
171 ev_edsam_finish = MPE_Log_get_event_number( );
172 ev_get_coords_start = MPE_Log_get_event_number( );
173 ev_get_coords_finish = MPE_Log_get_event_number( );
174 ev_ed_apply_cons_start = MPE_Log_get_event_number( );
175 ev_ed_apply_cons_finish = MPE_Log_get_event_number( );
176 ev_fit_to_reference_start = MPE_Log_get_event_number( );
177 ev_fit_to_reference_finish = MPE_Log_get_event_number( );
179 /* describe events: */
180 if (mpi_my_rank == 0)
183 MPE_Describe_state(ev_timestep1, ev_timestep2, "timestep START", "magenta" );
184 MPE_Describe_state(ev_force_start, ev_force_finish, "force", "cornflower blue" );
185 MPE_Describe_state(ev_do_fnbf_start, ev_do_fnbf_finish, "do_fnbf", "navy" );
186 MPE_Describe_state(ev_ns_start, ev_ns_finish, "neighbor search", "tomato" );
187 MPE_Describe_state(ev_calc_bonds_start, ev_calc_bonds_finish, "bonded forces", "slate blue" );
188 MPE_Describe_state(ev_global_stat_start, ev_global_stat_finish, "global stat", "firebrick3");
189 MPE_Describe_state(ev_update_fr_start, ev_update_fr_finish, "update forcerec", "goldenrod");
190 MPE_Describe_state(ev_clear_rvecs_start, ev_clear_rvecs_finish, "clear rvecs", "bisque");
191 MPE_Describe_state(ev_update_start, ev_update_finish, "update", "cornsilk");
192 MPE_Describe_state(ev_output_start, ev_output_finish, "output", "black");
193 MPE_Describe_state(ev_virial_start, ev_virial_finish, "calc_virial", "thistle4");
195 /* PME related events */
196 MPE_Describe_state(ev_pme_start, ev_pme_finish, "doing PME", "grey" );
197 MPE_Describe_state(ev_spread_on_grid_start, ev_spread_on_grid_finish, "spread", "dark orange" );
198 MPE_Describe_state(ev_sum_qgrid_start, ev_sum_qgrid_finish, "sum qgrid", "slate blue");
199 MPE_Describe_state(ev_gmxfft3d_start, ev_gmxfft3d_finish, "fft3d", "snow2" );
200 MPE_Describe_state(ev_solve_pme_start, ev_solve_pme_finish, "solve PME", "indian red" );
201 MPE_Describe_state(ev_gather_f_bsplines_start, ev_gather_f_bsplines_finish, "bsplines", "light sea green" );
202 MPE_Describe_state(ev_reduce_start, ev_reduce_finish, "reduce", "cyan1" );
203 MPE_Describe_state(ev_rscatter_start, ev_rscatter_finish, "rscatter", "cyan3" );
204 MPE_Describe_state(ev_alltoall_start, ev_alltoall_finish, "alltoall", "LightCyan4" );
205 MPE_Describe_state(ev_pmeredist_start, ev_pmeredist_finish, "pmeredist", "thistle" );
206 MPE_Describe_state(ev_init_pme_start, ev_init_pme_finish, "init PME", "snow4");
207 MPE_Describe_state(ev_send_coordinates_start, ev_send_coordinates_finish, "send_coordinates", "blue");
208 MPE_Describe_state(ev_sum_lrforces_start, ev_sum_lrforces_finish, "sum_LRforces", "lime green");
209 MPE_Describe_state(ev_sort_start, ev_sort_finish, "sort pme atoms", "brown");
210 MPE_Describe_state(ev_sum_qgrid_start, ev_sum_qgrid_finish, "sum charge grid", "medium orchid");
212 /* Shift related events */
213 MPE_Describe_state(ev_shift_start, ev_shift_finish, "shift", "orange");
214 MPE_Describe_state(ev_unshift_start, ev_unshift_finish, "unshift", "dark orange");
215 MPE_Describe_state(ev_mk_mshift_start, ev_mk_mshift_finish, "mk_mshift", "maroon");
217 /* Essential dynamics related events */
218 MPE_Describe_state(ev_edsam_start, ev_edsam_finish, "EDSAM", "deep sky blue");
219 MPE_Describe_state(ev_get_coords_start, ev_get_coords_finish, "ED get coords", "steel blue");
220 MPE_Describe_state(ev_ed_apply_cons_start, ev_ed_apply_cons_finish, "ED apply constr", "forest green");
221 MPE_Describe_state(ev_fit_to_reference_start, ev_fit_to_reference_finish, "ED fit to ref", "lavender");
230 fprintf(debug, "NNODES=%d, MYRANK=%d, HOSTNAME=%s\n",
231 mpi_num_nodes, mpi_my_rank, mpi_hostname);
235 *nnodes = mpi_num_nodes;
241 int gmx_node_num(void)
247 (void) MPI_Comm_size(MPI_COMM_WORLD, &i);
252 int gmx_node_rank(void)
258 (void) MPI_Comm_rank(MPI_COMM_WORLD, &i);
264 int gmx_hostname_num()
269 #ifdef GMX_THREAD_MPI
270 /* thread-MPI currently puts the thread number in the process name,
271 * we might want to change this, as this is inconsistent with what
272 * most MPI implementations would do when running on a single node.
276 int resultlen, hostnum, i, j;
277 char mpi_hostname[MPI_MAX_PROCESSOR_NAME], hostnum_str[MPI_MAX_PROCESSOR_NAME];
279 MPI_Get_processor_name(mpi_hostname, &resultlen);
280 /* This procedure can only differentiate nodes with host names
281 * that end on unique numbers.
285 /* Only parse the host name up to the first dot */
286 while (i < resultlen && mpi_hostname[i] != '.')
288 if (isdigit(mpi_hostname[i]))
290 hostnum_str[j++] = mpi_hostname[i];
294 hostnum_str[j] = '\0';
301 /* Use only the last 9 decimals, so we don't overflow an int */
302 hostnum = strtol(hostnum_str + max(0, j-9), NULL, 10);
307 fprintf(debug, "In gmx_setup_nodecomm: hostname '%s', hostnum %d\n",
308 mpi_hostname, hostnum);
315 void gmx_setup_nodecomm(FILE *fplog, t_commrec *cr)
318 int n, rank, hostnum, ng, ni;
320 /* Many MPI implementations do not optimize MPI_Allreduce
321 * (and probably also other global communication calls)
322 * for multi-core nodes connected by a network.
323 * We can optimize such communication by using one MPI call
324 * within each node and one between the nodes.
325 * For MVAPICH2 and Intel MPI this reduces the time for
326 * the global_stat communication by 25%
327 * for 2x2-core 3 GHz Woodcrest connected by mixed DDR/SDR Infiniband.
328 * B. Hess, November 2007
334 #ifndef GMX_THREAD_MPI
336 MPI_Comm_size(cr->mpi_comm_mygroup, &n);
337 MPI_Comm_rank(cr->mpi_comm_mygroup, &rank);
339 hostnum = gmx_hostname_num();
343 fprintf(debug, "In gmx_setup_nodecomm: splitting communicator of size %d\n", n);
347 /* The intra-node communicator, split on node number */
348 MPI_Comm_split(cr->mpi_comm_mygroup, hostnum, rank, &nc->comm_intra);
349 MPI_Comm_rank(nc->comm_intra, &nc->rank_intra);
352 fprintf(debug, "In gmx_setup_nodecomm: node rank %d rank_intra %d\n",
353 rank, nc->rank_intra);
355 /* The inter-node communicator, split on rank_intra.
356 * We actually only need the one for rank=0,
357 * but it is easier to create them all.
359 MPI_Comm_split(cr->mpi_comm_mygroup, nc->rank_intra, rank, &nc->comm_inter);
360 /* Check if this really created two step communication */
361 MPI_Comm_size(nc->comm_inter, &ng);
362 MPI_Comm_size(nc->comm_intra, &ni);
365 fprintf(debug, "In gmx_setup_nodecomm: groups %d, my group size %d\n",
369 if (getenv("GMX_NO_NODECOMM") == NULL &&
370 ((ng > 1 && ng < n) || (ni > 1 && ni < n)))
375 fprintf(fplog, "Using two step summing over %d groups of on average %.1f processes\n\n",
376 ng, (real)n/(real)ng);
378 if (nc->rank_intra > 0)
380 MPI_Comm_free(&nc->comm_inter);
385 /* One group or all processes in a separate group, use normal summing */
386 MPI_Comm_free(&nc->comm_inter);
387 MPI_Comm_free(&nc->comm_intra);
390 fprintf(debug, "In gmx_setup_nodecomm: not unsing separate inter- and intra-node communicators.\n");
395 /* tMPI runs only on a single node so just use the nodeid */
396 nc->rank_intra = cr->nodeid;
400 void gmx_init_intranode_counters(t_commrec *cr)
402 /* counters for PP+PME and PP-only processes on my physical node */
403 int nrank_intranode, rank_intranode;
404 int nrank_pp_intranode, rank_pp_intranode;
405 /* thread-MPI is not initialized when not running in parallel */
406 #if defined GMX_MPI && !defined GMX_THREAD_MPI
407 int nrank_world, rank_world;
408 int i, mynum, *num, *num_s, *num_pp, *num_pp_s;
410 MPI_Comm_size(MPI_COMM_WORLD, &nrank_world);
411 MPI_Comm_rank(MPI_COMM_WORLD, &rank_world);
413 /* Get the node number from the hostname to identify the nodes */
414 mynum = gmx_hostname_num();
416 /* We can't rely on MPI_IN_PLACE, so we need send and receive buffers */
417 snew(num, nrank_world);
418 snew(num_s, nrank_world);
419 snew(num_pp, nrank_world);
420 snew(num_pp_s, nrank_world);
422 num_s[rank_world] = mynum;
423 num_pp_s[rank_world] = (cr->duty & DUTY_PP) ? mynum : -1;
425 MPI_Allreduce(num_s, num, nrank_world, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
426 MPI_Allreduce(num_pp_s, num_pp, nrank_world, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
430 nrank_pp_intranode = 0;
431 rank_pp_intranode = 0;
432 for (i = 0; i < nrank_world; i++)
442 if ((cr->duty & DUTY_PP) && num_pp[i] == mynum)
444 nrank_pp_intranode++;
456 /* Serial or thread-MPI code: we run within a single physical node */
457 nrank_intranode = cr->nnodes;
458 rank_intranode = cr->sim_nodeid;
459 nrank_pp_intranode = cr->nnodes - cr->npmenodes;
460 rank_pp_intranode = cr->nodeid;
466 if (cr->duty & DUTY_PP && cr->duty & DUTY_PME)
468 sprintf(sbuf, "PP+PME");
472 sprintf(sbuf, "%s", cr->duty & DUTY_PP ? "PP" : "PME");
474 fprintf(debug, "On %3s node %d: nrank_intranode=%d, rank_intranode=%d, "
475 "nrank_pp_intranode=%d, rank_pp_intranode=%d\n",
476 sbuf, cr->sim_nodeid,
477 nrank_intranode, rank_intranode,
478 nrank_pp_intranode, rank_pp_intranode);
481 cr->nrank_intranode = nrank_intranode;
482 cr->rank_intranode = rank_intranode;
483 cr->nrank_pp_intranode = nrank_pp_intranode;
484 cr->rank_pp_intranode = rank_pp_intranode;
488 void gmx_barrier(const t_commrec *cr)
491 gmx_call("gmx_barrier");
493 MPI_Barrier(cr->mpi_comm_mygroup);
497 void gmx_abort(int noderank, int nnodes, int errorno)
500 gmx_call("gmx_abort");
502 #ifdef GMX_THREAD_MPI
503 fprintf(stderr, "Halting program %s\n", ShortProgram());
509 fprintf(stderr, "Halting parallel program %s on CPU %d out of %d\n",
510 ShortProgram(), noderank, nnodes);
514 fprintf(stderr, "Halting program %s\n", ShortProgram());
518 MPI_Abort(MPI_COMM_WORLD, errorno);
524 void gmx_bcast(int nbytes, void *b, const t_commrec *cr)
527 gmx_call("gmx_bast");
529 MPI_Bcast(b, nbytes, MPI_BYTE, MASTERRANK(cr), cr->mpi_comm_mygroup);
533 void gmx_bcast_sim(int nbytes, void *b, const t_commrec *cr)
536 gmx_call("gmx_bast");
538 MPI_Bcast(b, nbytes, MPI_BYTE, MASTERRANK(cr), cr->mpi_comm_mysim);
542 void gmx_sumd(int nr, double r[], const t_commrec *cr)
545 gmx_call("gmx_sumd");
547 #if defined(MPI_IN_PLACE_EXISTS) || defined(GMX_THREAD_MPI)
550 if (cr->nc.rank_intra == 0)
552 /* Use two step summing. */
553 MPI_Reduce(MPI_IN_PLACE, r, nr, MPI_DOUBLE, MPI_SUM, 0,
555 /* Sum the roots of the internal (intra) buffers. */
556 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_DOUBLE, MPI_SUM,
561 /* This is here because of the silly MPI specification
562 that MPI_IN_PLACE should be put in sendbuf instead of recvbuf */
563 MPI_Reduce(r, NULL, nr, MPI_DOUBLE, MPI_SUM, 0, cr->nc.comm_intra);
565 MPI_Bcast(r, nr, MPI_DOUBLE, 0, cr->nc.comm_intra);
569 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_DOUBLE, MPI_SUM,
570 cr->mpi_comm_mygroup);
575 if (nr > cr->mpb->dbuf_alloc)
577 cr->mpb->dbuf_alloc = nr;
578 srenew(cr->mpb->dbuf, cr->mpb->dbuf_alloc);
582 /* Use two step summing */
583 MPI_Allreduce(r, cr->mpb->dbuf, nr, MPI_DOUBLE, MPI_SUM, cr->nc.comm_intra);
584 if (cr->nc.rank_intra == 0)
586 /* Sum with the buffers reversed */
587 MPI_Allreduce(cr->mpb->dbuf, r, nr, MPI_DOUBLE, MPI_SUM,
590 MPI_Bcast(r, nr, MPI_DOUBLE, 0, cr->nc.comm_intra);
594 MPI_Allreduce(r, cr->mpb->dbuf, nr, MPI_DOUBLE, MPI_SUM,
595 cr->mpi_comm_mygroup);
596 for (i = 0; i < nr; i++)
598 r[i] = cr->mpb->dbuf[i];
605 void gmx_sumf(int nr, float r[], const t_commrec *cr)
608 gmx_call("gmx_sumf");
610 #if defined(MPI_IN_PLACE_EXISTS) || defined(GMX_THREAD_MPI)
613 /* Use two step summing. */
614 if (cr->nc.rank_intra == 0)
616 MPI_Reduce(MPI_IN_PLACE, r, nr, MPI_FLOAT, MPI_SUM, 0,
618 /* Sum the roots of the internal (intra) buffers */
619 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_FLOAT, MPI_SUM,
624 /* This is here because of the silly MPI specification
625 that MPI_IN_PLACE should be put in sendbuf instead of recvbuf */
626 MPI_Reduce(r, NULL, nr, MPI_FLOAT, MPI_SUM, 0, cr->nc.comm_intra);
628 MPI_Bcast(r, nr, MPI_FLOAT, 0, cr->nc.comm_intra);
632 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_FLOAT, MPI_SUM, cr->mpi_comm_mygroup);
637 if (nr > cr->mpb->fbuf_alloc)
639 cr->mpb->fbuf_alloc = nr;
640 srenew(cr->mpb->fbuf, cr->mpb->fbuf_alloc);
644 /* Use two step summing */
645 MPI_Allreduce(r, cr->mpb->fbuf, nr, MPI_FLOAT, MPI_SUM, cr->nc.comm_intra);
646 if (cr->nc.rank_intra == 0)
648 /* Sum with the buffers reversed */
649 MPI_Allreduce(cr->mpb->fbuf, r, nr, MPI_FLOAT, MPI_SUM,
652 MPI_Bcast(r, nr, MPI_FLOAT, 0, cr->nc.comm_intra);
656 MPI_Allreduce(r, cr->mpb->fbuf, nr, MPI_FLOAT, MPI_SUM,
657 cr->mpi_comm_mygroup);
658 for (i = 0; i < nr; i++)
660 r[i] = cr->mpb->fbuf[i];
667 void gmx_sumi(int nr, int r[], const t_commrec *cr)
670 gmx_call("gmx_sumi");
672 #if defined(MPI_IN_PLACE_EXISTS) || defined(GMX_THREAD_MPI)
675 /* Use two step summing */
676 if (cr->nc.rank_intra == 0)
678 MPI_Reduce(MPI_IN_PLACE, r, nr, MPI_INT, MPI_SUM, 0, cr->nc.comm_intra);
679 /* Sum with the buffers reversed */
680 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_INT, MPI_SUM, cr->nc.comm_inter);
684 /* This is here because of the silly MPI specification
685 that MPI_IN_PLACE should be put in sendbuf instead of recvbuf */
686 MPI_Reduce(r, NULL, nr, MPI_INT, MPI_SUM, 0, cr->nc.comm_intra);
688 MPI_Bcast(r, nr, MPI_INT, 0, cr->nc.comm_intra);
692 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_INT, MPI_SUM, cr->mpi_comm_mygroup);
697 if (nr > cr->mpb->ibuf_alloc)
699 cr->mpb->ibuf_alloc = nr;
700 srenew(cr->mpb->ibuf, cr->mpb->ibuf_alloc);
704 /* Use two step summing */
705 MPI_Allreduce(r, cr->mpb->ibuf, nr, MPI_INT, MPI_SUM, cr->nc.comm_intra);
706 if (cr->nc.rank_intra == 0)
708 /* Sum with the buffers reversed */
709 MPI_Allreduce(cr->mpb->ibuf, r, nr, MPI_INT, MPI_SUM, cr->nc.comm_inter);
711 MPI_Bcast(r, nr, MPI_INT, 0, cr->nc.comm_intra);
715 MPI_Allreduce(r, cr->mpb->ibuf, nr, MPI_INT, MPI_SUM, cr->mpi_comm_mygroup);
716 for (i = 0; i < nr; i++)
718 r[i] = cr->mpb->ibuf[i];
725 void gmx_sumli(int nr, gmx_large_int_t r[], const t_commrec *cr)
728 gmx_call("gmx_sumli");
730 #if defined(MPI_IN_PLACE_EXISTS) || defined(GMX_THREAD_MPI)
733 /* Use two step summing */
734 if (cr->nc.rank_intra == 0)
736 MPI_Reduce(MPI_IN_PLACE, r, nr, GMX_MPI_LARGE_INT, MPI_SUM, 0,
738 /* Sum with the buffers reversed */
739 MPI_Allreduce(MPI_IN_PLACE, r, nr, GMX_MPI_LARGE_INT, MPI_SUM,
744 /* This is here because of the silly MPI specification
745 that MPI_IN_PLACE should be put in sendbuf instead of recvbuf */
746 MPI_Reduce(r, NULL, nr, GMX_MPI_LARGE_INT, MPI_SUM, 0, cr->nc.comm_intra);
748 MPI_Bcast(r, nr, GMX_MPI_LARGE_INT, 0, cr->nc.comm_intra);
752 MPI_Allreduce(MPI_IN_PLACE, r, nr, GMX_MPI_LARGE_INT, MPI_SUM, cr->mpi_comm_mygroup);
757 if (nr > cr->mpb->libuf_alloc)
759 cr->mpb->libuf_alloc = nr;
760 srenew(cr->mpb->libuf, cr->mpb->libuf_alloc);
764 /* Use two step summing */
765 MPI_Allreduce(r, cr->mpb->libuf, nr, GMX_MPI_LARGE_INT, MPI_SUM,
767 if (cr->nc.rank_intra == 0)
769 /* Sum with the buffers reversed */
770 MPI_Allreduce(cr->mpb->libuf, r, nr, GMX_MPI_LARGE_INT, MPI_SUM,
773 MPI_Bcast(r, nr, GMX_MPI_LARGE_INT, 0, cr->nc.comm_intra);
777 MPI_Allreduce(r, cr->mpb->libuf, nr, GMX_MPI_LARGE_INT, MPI_SUM,
778 cr->mpi_comm_mygroup);
779 for (i = 0; i < nr; i++)
781 r[i] = cr->mpb->libuf[i];
791 void gmx_sumd_comm(int nr, double r[], MPI_Comm mpi_comm)
793 #if defined(MPI_IN_PLACE_EXISTS) || defined(GMX_THREAD_MPI)
794 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_DOUBLE, MPI_SUM, mpi_comm);
796 /* this function is only used in code that is not performance critical,
797 (during setup, when comm_rec is not the appropriate communication
798 structure), so this isn't as bad as it looks. */
803 MPI_Allreduce(r, buf, nr, MPI_DOUBLE, MPI_SUM, mpi_comm);
804 for (i = 0; i < nr; i++)
814 void gmx_sumf_comm(int nr, float r[], MPI_Comm mpi_comm)
816 #if defined(MPI_IN_PLACE_EXISTS) || defined(GMX_THREAD_MPI)
817 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_FLOAT, MPI_SUM, mpi_comm);
819 /* this function is only used in code that is not performance critical,
820 (during setup, when comm_rec is not the appropriate communication
821 structure), so this isn't as bad as it looks. */
826 MPI_Allreduce(r, buf, nr, MPI_FLOAT, MPI_SUM, mpi_comm);
827 for (i = 0; i < nr; i++)
836 void gmx_sumd_sim(int nr, double r[], const gmx_multisim_t *ms)
839 gmx_call("gmx_sumd_sim");
841 gmx_sumd_comm(nr, r, ms->mpi_comm_masters);
845 void gmx_sumf_sim(int nr, float r[], const gmx_multisim_t *ms)
848 gmx_call("gmx_sumf_sim");
850 gmx_sumf_comm(nr, r, ms->mpi_comm_masters);
854 void gmx_sumi_sim(int nr, int r[], const gmx_multisim_t *ms)
857 gmx_call("gmx_sumi_sim");
859 #if defined(MPI_IN_PLACE_EXISTS) || defined(GMX_THREAD_MPI)
860 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_INT, MPI_SUM, ms->mpi_comm_masters);
862 /* this is thread-unsafe, but it will do for now: */
865 if (nr > ms->mpb->ibuf_alloc)
867 ms->mpb->ibuf_alloc = nr;
868 srenew(ms->mpb->ibuf, ms->mpb->ibuf_alloc);
870 MPI_Allreduce(r, ms->mpb->ibuf, nr, MPI_INT, MPI_SUM, ms->mpi_comm_masters);
871 for (i = 0; i < nr; i++)
873 r[i] = ms->mpb->ibuf[i];
879 void gmx_sumli_sim(int nr, gmx_large_int_t r[], const gmx_multisim_t *ms)
882 gmx_call("gmx_sumli_sim");
884 #if defined(MPI_IN_PLACE_EXISTS) || defined(GMX_THREAD_MPI)
885 MPI_Allreduce(MPI_IN_PLACE, r, nr, GMX_MPI_LARGE_INT, MPI_SUM,
886 ms->mpi_comm_masters);
888 /* this is thread-unsafe, but it will do for now: */
891 if (nr > ms->mpb->libuf_alloc)
893 ms->mpb->libuf_alloc = nr;
894 srenew(ms->mpb->libuf, ms->mpb->libuf_alloc);
896 MPI_Allreduce(r, ms->mpb->libuf, nr, GMX_MPI_LARGE_INT, MPI_SUM,
897 ms->mpi_comm_masters);
898 for (i = 0; i < nr; i++)
900 r[i] = ms->mpb->libuf[i];
907 void gmx_finalize_par(void)
910 /* Compiled without MPI, no MPI finalizing needed */
913 int initialized, finalized;
916 MPI_Initialized(&initialized);
921 /* just as a check; we don't want to finalize twice */
922 MPI_Finalized(&finalized);
928 /* We sync the processes here to try to avoid problems
929 * with buggy MPI implementations that could cause
930 * unfinished processes to terminate.
932 MPI_Barrier(MPI_COMM_WORLD);
935 if (DOMAINDECOMP(cr)) {
936 if (cr->npmenodes > 0 || cr->dd->bCartesian)
937 MPI_Comm_free(&cr->mpi_comm_mygroup);
938 if (cr->dd->bCartesian)
939 MPI_Comm_free(&cr->mpi_comm_mysim);
943 /* Apparently certain mpich implementations cause problems
944 * with MPI_Finalize. In that case comment out MPI_Finalize.
948 fprintf(debug, "Will call MPI_Finalize now\n");
951 ret = MPI_Finalize();
954 fprintf(debug, "Return code from MPI_Finalize = %d\n", ret);