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"
61 #include "mpelogging.h"
63 /* The source code in this file should be thread-safe.
64 Please keep it that way. */
66 gmx_bool gmx_mpi_initialized(void)
78 int gmx_setup(int *argc, char **argv, int *nnodes)
81 gmx_call("gmx_setup");
85 int resultlen; /* actual length of node name */
89 char mpi_hostname[MPI_MAX_PROCESSOR_NAME];
91 /* Call the MPI routines */
94 (void) fah_MPI_Init(argc, &argv);
96 (void) MPI_Init(argc, &argv);
99 (void) MPI_Comm_size( MPI_COMM_WORLD, &mpi_num_nodes );
100 (void) MPI_Comm_rank( MPI_COMM_WORLD, &mpi_my_rank );
101 (void) MPI_Get_processor_name( mpi_hostname, &resultlen );
105 /* MPE logging routines. Get event IDs from MPE: */
107 ev_timestep1 = MPE_Log_get_event_number( );
108 ev_timestep2 = MPE_Log_get_event_number( );
109 ev_force_start = MPE_Log_get_event_number( );
110 ev_force_finish = MPE_Log_get_event_number( );
111 ev_do_fnbf_start = MPE_Log_get_event_number( );
112 ev_do_fnbf_finish = MPE_Log_get_event_number( );
113 ev_ns_start = MPE_Log_get_event_number( );
114 ev_ns_finish = MPE_Log_get_event_number( );
115 ev_calc_bonds_start = MPE_Log_get_event_number( );
116 ev_calc_bonds_finish = MPE_Log_get_event_number( );
117 ev_global_stat_start = MPE_Log_get_event_number( );
118 ev_global_stat_finish = MPE_Log_get_event_number( );
119 ev_virial_start = MPE_Log_get_event_number( );
120 ev_virial_finish = MPE_Log_get_event_number( );
122 /* Shift related events */
123 ev_shift_start = MPE_Log_get_event_number( );
124 ev_shift_finish = MPE_Log_get_event_number( );
125 ev_unshift_start = MPE_Log_get_event_number( );
126 ev_unshift_finish = MPE_Log_get_event_number( );
127 ev_mk_mshift_start = MPE_Log_get_event_number( );
128 ev_mk_mshift_finish = MPE_Log_get_event_number( );
130 /* PME related events */
131 ev_pme_start = MPE_Log_get_event_number( );
132 ev_pme_finish = MPE_Log_get_event_number( );
133 ev_spread_on_grid_start = MPE_Log_get_event_number( );
134 ev_spread_on_grid_finish = MPE_Log_get_event_number( );
135 ev_sum_qgrid_start = MPE_Log_get_event_number( );
136 ev_sum_qgrid_finish = MPE_Log_get_event_number( );
137 ev_gmxfft3d_start = MPE_Log_get_event_number( );
138 ev_gmxfft3d_finish = MPE_Log_get_event_number( );
139 ev_solve_pme_start = MPE_Log_get_event_number( );
140 ev_solve_pme_finish = MPE_Log_get_event_number( );
141 ev_gather_f_bsplines_start = MPE_Log_get_event_number( );
142 ev_gather_f_bsplines_finish = MPE_Log_get_event_number( );
143 ev_reduce_start = MPE_Log_get_event_number( );
144 ev_reduce_finish = MPE_Log_get_event_number( );
145 ev_rscatter_start = MPE_Log_get_event_number( );
146 ev_rscatter_finish = MPE_Log_get_event_number( );
147 ev_alltoall_start = MPE_Log_get_event_number( );
148 ev_alltoall_finish = MPE_Log_get_event_number( );
149 ev_pmeredist_start = MPE_Log_get_event_number( );
150 ev_pmeredist_finish = MPE_Log_get_event_number( );
151 ev_init_pme_start = MPE_Log_get_event_number( );
152 ev_init_pme_finish = MPE_Log_get_event_number( );
153 ev_send_coordinates_start = MPE_Log_get_event_number( );
154 ev_send_coordinates_finish = MPE_Log_get_event_number( );
155 ev_update_fr_start = MPE_Log_get_event_number( );
156 ev_update_fr_finish = MPE_Log_get_event_number( );
157 ev_clear_rvecs_start = MPE_Log_get_event_number( );
158 ev_clear_rvecs_finish = MPE_Log_get_event_number( );
159 ev_update_start = MPE_Log_get_event_number( );
160 ev_update_finish = MPE_Log_get_event_number( );
161 ev_output_start = MPE_Log_get_event_number( );
162 ev_output_finish = MPE_Log_get_event_number( );
163 ev_sum_lrforces_start = MPE_Log_get_event_number( );
164 ev_sum_lrforces_finish = MPE_Log_get_event_number( );
165 ev_sort_start = MPE_Log_get_event_number( );
166 ev_sort_finish = MPE_Log_get_event_number( );
167 ev_sum_qgrid_start = MPE_Log_get_event_number( );
168 ev_sum_qgrid_finish = MPE_Log_get_event_number( );
170 /* Essential dynamics related events */
171 ev_edsam_start = MPE_Log_get_event_number( );
172 ev_edsam_finish = MPE_Log_get_event_number( );
173 ev_get_coords_start = MPE_Log_get_event_number( );
174 ev_get_coords_finish = MPE_Log_get_event_number( );
175 ev_ed_apply_cons_start = MPE_Log_get_event_number( );
176 ev_ed_apply_cons_finish = MPE_Log_get_event_number( );
177 ev_fit_to_reference_start = MPE_Log_get_event_number( );
178 ev_fit_to_reference_finish = MPE_Log_get_event_number( );
180 /* describe events: */
181 if (mpi_my_rank == 0)
184 MPE_Describe_state(ev_timestep1, ev_timestep2, "timestep START", "magenta" );
185 MPE_Describe_state(ev_force_start, ev_force_finish, "force", "cornflower blue" );
186 MPE_Describe_state(ev_do_fnbf_start, ev_do_fnbf_finish, "do_fnbf", "navy" );
187 MPE_Describe_state(ev_ns_start, ev_ns_finish, "neighbor search", "tomato" );
188 MPE_Describe_state(ev_calc_bonds_start, ev_calc_bonds_finish, "bonded forces", "slate blue" );
189 MPE_Describe_state(ev_global_stat_start, ev_global_stat_finish, "global stat", "firebrick3");
190 MPE_Describe_state(ev_update_fr_start, ev_update_fr_finish, "update forcerec", "goldenrod");
191 MPE_Describe_state(ev_clear_rvecs_start, ev_clear_rvecs_finish, "clear rvecs", "bisque");
192 MPE_Describe_state(ev_update_start, ev_update_finish, "update", "cornsilk");
193 MPE_Describe_state(ev_output_start, ev_output_finish, "output", "black");
194 MPE_Describe_state(ev_virial_start, ev_virial_finish, "calc_virial", "thistle4");
196 /* PME related events */
197 MPE_Describe_state(ev_pme_start, ev_pme_finish, "doing PME", "grey" );
198 MPE_Describe_state(ev_spread_on_grid_start, ev_spread_on_grid_finish, "spread", "dark orange" );
199 MPE_Describe_state(ev_sum_qgrid_start, ev_sum_qgrid_finish, "sum qgrid", "slate blue");
200 MPE_Describe_state(ev_gmxfft3d_start, ev_gmxfft3d_finish, "fft3d", "snow2" );
201 MPE_Describe_state(ev_solve_pme_start, ev_solve_pme_finish, "solve PME", "indian red" );
202 MPE_Describe_state(ev_gather_f_bsplines_start, ev_gather_f_bsplines_finish, "bsplines", "light sea green" );
203 MPE_Describe_state(ev_reduce_start, ev_reduce_finish, "reduce", "cyan1" );
204 MPE_Describe_state(ev_rscatter_start, ev_rscatter_finish, "rscatter", "cyan3" );
205 MPE_Describe_state(ev_alltoall_start, ev_alltoall_finish, "alltoall", "LightCyan4" );
206 MPE_Describe_state(ev_pmeredist_start, ev_pmeredist_finish, "pmeredist", "thistle" );
207 MPE_Describe_state(ev_init_pme_start, ev_init_pme_finish, "init PME", "snow4");
208 MPE_Describe_state(ev_send_coordinates_start, ev_send_coordinates_finish, "send_coordinates", "blue");
209 MPE_Describe_state(ev_sum_lrforces_start, ev_sum_lrforces_finish, "sum_LRforces", "lime green");
210 MPE_Describe_state(ev_sort_start, ev_sort_finish, "sort pme atoms", "brown");
211 MPE_Describe_state(ev_sum_qgrid_start, ev_sum_qgrid_finish, "sum charge grid", "medium orchid");
213 /* Shift related events */
214 MPE_Describe_state(ev_shift_start, ev_shift_finish, "shift", "orange");
215 MPE_Describe_state(ev_unshift_start, ev_unshift_finish, "unshift", "dark orange");
216 MPE_Describe_state(ev_mk_mshift_start, ev_mk_mshift_finish, "mk_mshift", "maroon");
218 /* Essential dynamics related events */
219 MPE_Describe_state(ev_edsam_start, ev_edsam_finish, "EDSAM", "deep sky blue");
220 MPE_Describe_state(ev_get_coords_start, ev_get_coords_finish, "ED get coords", "steel blue");
221 MPE_Describe_state(ev_ed_apply_cons_start, ev_ed_apply_cons_finish, "ED apply constr", "forest green");
222 MPE_Describe_state(ev_fit_to_reference_start, ev_fit_to_reference_finish, "ED fit to ref", "lavender");
231 fprintf(debug, "NNODES=%d, MYRANK=%d, HOSTNAME=%s\n",
232 mpi_num_nodes, mpi_my_rank, mpi_hostname);
236 *nnodes = mpi_num_nodes;
242 int gmx_node_num(void)
248 (void) MPI_Comm_size(MPI_COMM_WORLD, &i);
253 int gmx_node_rank(void)
259 (void) MPI_Comm_rank(MPI_COMM_WORLD, &i);
264 #if defined GMX_LIB_MPI && defined GMX_TARGET_BGQ
265 #include <spi/include/kernel/location.h>
268 int gmx_physicalnode_id_hash(void)
273 /* We have a single physical node */
277 char mpi_hostname[MPI_MAX_PROCESSOR_NAME];
279 /* This procedure can only differentiate nodes with different names.
280 * Architectures where different physical nodes have identical names,
281 * such as IBM Blue Gene, should use an architecture specific solution.
283 MPI_Get_processor_name(mpi_hostname, &resultlen);
285 /* The string hash function returns an unsigned int. We cast to an int.
286 * Negative numbers are converted to positive by setting the sign bit to 0.
287 * This makes the hash one bit smaller.
288 * A 63-bit hash (with 64-bit int) should be enough for unique node hashes,
289 * even on a million node machine. 31 bits might not be enough though!
292 (int)gmx_string_fullhash_func(mpi_hostname, gmx_string_hash_init);
302 /* TODO: this function should be fully replaced by gmx_physicalnode_id_hash */
303 int gmx_hostname_num()
308 #ifdef GMX_THREAD_MPI
309 /* thread-MPI currently puts the thread number in the process name,
310 * we might want to change this, as this is inconsistent with what
311 * most MPI implementations would do when running on a single node.
315 int resultlen, hostnum, i, j;
316 char mpi_hostname[MPI_MAX_PROCESSOR_NAME], hostnum_str[MPI_MAX_PROCESSOR_NAME];
318 MPI_Get_processor_name(mpi_hostname, &resultlen);
319 #ifdef GMX_TARGET_BGQ
320 Personality_t personality;
321 Kernel_GetPersonality(&personality, sizeof(personality));
322 /* Each MPI rank has a unique coordinate in a 6-dimensional space
323 (A,B,C,D,E,T), with dimensions A-E corresponding to different
324 physical nodes, and T within each node. Each node has sixteen
325 physical cores, each of which can have up to four hardware
326 threads, so 0 <= T <= 63 (but the maximum value of T depends on
327 the confituration of ranks and OpenMP threads per
328 node). However, T is irrelevant for computing a suitable return
329 value for gmx_hostname_num().
331 hostnum = personality.Network_Config.Acoord;
332 hostnum *= personality.Network_Config.Bnodes;
333 hostnum += personality.Network_Config.Bcoord;
334 hostnum *= personality.Network_Config.Cnodes;
335 hostnum += personality.Network_Config.Ccoord;
336 hostnum *= personality.Network_Config.Dnodes;
337 hostnum += personality.Network_Config.Dcoord;
338 hostnum *= personality.Network_Config.Enodes;
339 hostnum += personality.Network_Config.Ecoord;
341 /* This procedure can only differentiate nodes with host names
342 * that end on unique numbers.
346 /* Only parse the host name up to the first dot */
347 while (i < resultlen && mpi_hostname[i] != '.')
349 if (isdigit(mpi_hostname[i]))
351 hostnum_str[j++] = mpi_hostname[i];
355 hostnum_str[j] = '\0';
362 /* Use only the last 9 decimals, so we don't overflow an int */
363 hostnum = strtol(hostnum_str + max(0, j-9), NULL, 10);
369 fprintf(debug, "In gmx_hostname_num: hostname '%s', hostnum %d\n",
370 mpi_hostname, hostnum);
371 #ifdef GMX_TARGET_BGQ
373 "Torus ID A: %d / %d B: %d / %d C: %d / %d D: %d / %d E: %d / %d\nNode ID T: %d / %d core: %d / %d hardware thread: %d / %d\n",
374 personality.Network_Config.Acoord,
375 personality.Network_Config.Anodes,
376 personality.Network_Config.Bcoord,
377 personality.Network_Config.Bnodes,
378 personality.Network_Config.Ccoord,
379 personality.Network_Config.Cnodes,
380 personality.Network_Config.Dcoord,
381 personality.Network_Config.Dnodes,
382 personality.Network_Config.Ecoord,
383 personality.Network_Config.Enodes,
384 Kernel_ProcessorCoreID(),
386 Kernel_ProcessorID(),
388 Kernel_ProcessorThreadID(),
397 void gmx_setup_nodecomm(FILE *fplog, t_commrec *cr)
400 int n, rank, hostnum, ng, ni;
402 /* Many MPI implementations do not optimize MPI_Allreduce
403 * (and probably also other global communication calls)
404 * for multi-core nodes connected by a network.
405 * We can optimize such communication by using one MPI call
406 * within each node and one between the nodes.
407 * For MVAPICH2 and Intel MPI this reduces the time for
408 * the global_stat communication by 25%
409 * for 2x2-core 3 GHz Woodcrest connected by mixed DDR/SDR Infiniband.
410 * B. Hess, November 2007
416 #ifndef GMX_THREAD_MPI
418 MPI_Comm_size(cr->mpi_comm_mygroup, &n);
419 MPI_Comm_rank(cr->mpi_comm_mygroup, &rank);
421 hostnum = gmx_hostname_num();
425 fprintf(debug, "In gmx_setup_nodecomm: splitting communicator of size %d\n", n);
429 /* The intra-node communicator, split on node number */
430 MPI_Comm_split(cr->mpi_comm_mygroup, hostnum, rank, &nc->comm_intra);
431 MPI_Comm_rank(nc->comm_intra, &nc->rank_intra);
434 fprintf(debug, "In gmx_setup_nodecomm: node rank %d rank_intra %d\n",
435 rank, nc->rank_intra);
437 /* The inter-node communicator, split on rank_intra.
438 * We actually only need the one for rank=0,
439 * but it is easier to create them all.
441 MPI_Comm_split(cr->mpi_comm_mygroup, nc->rank_intra, rank, &nc->comm_inter);
442 /* Check if this really created two step communication */
443 MPI_Comm_size(nc->comm_inter, &ng);
444 MPI_Comm_size(nc->comm_intra, &ni);
447 fprintf(debug, "In gmx_setup_nodecomm: groups %d, my group size %d\n",
451 if (getenv("GMX_NO_NODECOMM") == NULL &&
452 ((ng > 1 && ng < n) || (ni > 1 && ni < n)))
457 fprintf(fplog, "Using two step summing over %d groups of on average %.1f processes\n\n",
458 ng, (real)n/(real)ng);
460 if (nc->rank_intra > 0)
462 MPI_Comm_free(&nc->comm_inter);
467 /* One group or all processes in a separate group, use normal summing */
468 MPI_Comm_free(&nc->comm_inter);
469 MPI_Comm_free(&nc->comm_intra);
472 fprintf(debug, "In gmx_setup_nodecomm: not unsing separate inter- and intra-node communicators.\n");
477 /* tMPI runs only on a single node so just use the nodeid */
478 nc->rank_intra = cr->nodeid;
482 void gmx_init_intranode_counters(t_commrec *cr)
484 /* counters for PP+PME and PP-only processes on my physical node */
485 int nrank_intranode, rank_intranode;
486 int nrank_pp_intranode, rank_pp_intranode;
487 /* thread-MPI is not initialized when not running in parallel */
488 #if defined GMX_MPI && !defined GMX_THREAD_MPI
489 int nrank_world, rank_world;
490 int i, mynum, *num, *num_s, *num_pp, *num_pp_s;
492 MPI_Comm_size(MPI_COMM_WORLD, &nrank_world);
493 MPI_Comm_rank(MPI_COMM_WORLD, &rank_world);
495 /* Get the node number from the hostname to identify the nodes */
496 mynum = gmx_hostname_num();
498 /* We can't rely on MPI_IN_PLACE, so we need send and receive buffers */
499 snew(num, nrank_world);
500 snew(num_s, nrank_world);
501 snew(num_pp, nrank_world);
502 snew(num_pp_s, nrank_world);
504 num_s[rank_world] = mynum;
505 num_pp_s[rank_world] = (cr->duty & DUTY_PP) ? mynum : -1;
507 MPI_Allreduce(num_s, num, nrank_world, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
508 MPI_Allreduce(num_pp_s, num_pp, nrank_world, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
512 nrank_pp_intranode = 0;
513 rank_pp_intranode = 0;
514 for (i = 0; i < nrank_world; i++)
524 if ((cr->duty & DUTY_PP) && num_pp[i] == mynum)
526 nrank_pp_intranode++;
538 /* Serial or thread-MPI code: we run within a single physical node */
539 nrank_intranode = cr->nnodes;
540 rank_intranode = cr->sim_nodeid;
541 nrank_pp_intranode = cr->nnodes - cr->npmenodes;
542 rank_pp_intranode = cr->nodeid;
548 if (cr->duty & DUTY_PP && cr->duty & DUTY_PME)
550 sprintf(sbuf, "PP+PME");
554 sprintf(sbuf, "%s", cr->duty & DUTY_PP ? "PP" : "PME");
556 fprintf(debug, "On %3s node %d: nrank_intranode=%d, rank_intranode=%d, "
557 "nrank_pp_intranode=%d, rank_pp_intranode=%d\n",
558 sbuf, cr->sim_nodeid,
559 nrank_intranode, rank_intranode,
560 nrank_pp_intranode, rank_pp_intranode);
563 cr->nrank_intranode = nrank_intranode;
564 cr->rank_intranode = rank_intranode;
565 cr->nrank_pp_intranode = nrank_pp_intranode;
566 cr->rank_pp_intranode = rank_pp_intranode;
570 void gmx_barrier(const t_commrec *cr)
573 gmx_call("gmx_barrier");
575 MPI_Barrier(cr->mpi_comm_mygroup);
579 void gmx_abort(int noderank, int nnodes, int errorno)
582 gmx_call("gmx_abort");
584 #ifdef GMX_THREAD_MPI
585 fprintf(stderr, "Halting program %s\n", ShortProgram());
591 fprintf(stderr, "Halting parallel program %s on CPU %d out of %d\n",
592 ShortProgram(), noderank, nnodes);
596 fprintf(stderr, "Halting program %s\n", ShortProgram());
600 MPI_Abort(MPI_COMM_WORLD, errorno);
606 void gmx_bcast(int nbytes, void *b, const t_commrec *cr)
609 gmx_call("gmx_bast");
611 MPI_Bcast(b, nbytes, MPI_BYTE, MASTERRANK(cr), cr->mpi_comm_mygroup);
615 void gmx_bcast_sim(int nbytes, void *b, const t_commrec *cr)
618 gmx_call("gmx_bast");
620 MPI_Bcast(b, nbytes, MPI_BYTE, MASTERRANK(cr), cr->mpi_comm_mysim);
624 void gmx_sumd(int nr, double r[], const t_commrec *cr)
627 gmx_call("gmx_sumd");
629 #if defined(MPI_IN_PLACE_EXISTS) || defined(GMX_THREAD_MPI)
632 if (cr->nc.rank_intra == 0)
634 /* Use two step summing. */
635 MPI_Reduce(MPI_IN_PLACE, r, nr, MPI_DOUBLE, MPI_SUM, 0,
637 /* Sum the roots of the internal (intra) buffers. */
638 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_DOUBLE, MPI_SUM,
643 /* This is here because of the silly MPI specification
644 that MPI_IN_PLACE should be put in sendbuf instead of recvbuf */
645 MPI_Reduce(r, NULL, nr, MPI_DOUBLE, MPI_SUM, 0, cr->nc.comm_intra);
647 MPI_Bcast(r, nr, MPI_DOUBLE, 0, cr->nc.comm_intra);
651 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_DOUBLE, MPI_SUM,
652 cr->mpi_comm_mygroup);
657 if (nr > cr->mpb->dbuf_alloc)
659 cr->mpb->dbuf_alloc = nr;
660 srenew(cr->mpb->dbuf, cr->mpb->dbuf_alloc);
664 /* Use two step summing */
665 MPI_Allreduce(r, cr->mpb->dbuf, nr, MPI_DOUBLE, MPI_SUM, cr->nc.comm_intra);
666 if (cr->nc.rank_intra == 0)
668 /* Sum with the buffers reversed */
669 MPI_Allreduce(cr->mpb->dbuf, r, nr, MPI_DOUBLE, MPI_SUM,
672 MPI_Bcast(r, nr, MPI_DOUBLE, 0, cr->nc.comm_intra);
676 MPI_Allreduce(r, cr->mpb->dbuf, nr, MPI_DOUBLE, MPI_SUM,
677 cr->mpi_comm_mygroup);
678 for (i = 0; i < nr; i++)
680 r[i] = cr->mpb->dbuf[i];
687 void gmx_sumf(int nr, float r[], const t_commrec *cr)
690 gmx_call("gmx_sumf");
692 #if defined(MPI_IN_PLACE_EXISTS) || defined(GMX_THREAD_MPI)
695 /* Use two step summing. */
696 if (cr->nc.rank_intra == 0)
698 MPI_Reduce(MPI_IN_PLACE, r, nr, MPI_FLOAT, MPI_SUM, 0,
700 /* Sum the roots of the internal (intra) buffers */
701 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_FLOAT, MPI_SUM,
706 /* This is here because of the silly MPI specification
707 that MPI_IN_PLACE should be put in sendbuf instead of recvbuf */
708 MPI_Reduce(r, NULL, nr, MPI_FLOAT, MPI_SUM, 0, cr->nc.comm_intra);
710 MPI_Bcast(r, nr, MPI_FLOAT, 0, cr->nc.comm_intra);
714 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_FLOAT, MPI_SUM, cr->mpi_comm_mygroup);
719 if (nr > cr->mpb->fbuf_alloc)
721 cr->mpb->fbuf_alloc = nr;
722 srenew(cr->mpb->fbuf, cr->mpb->fbuf_alloc);
726 /* Use two step summing */
727 MPI_Allreduce(r, cr->mpb->fbuf, nr, MPI_FLOAT, MPI_SUM, cr->nc.comm_intra);
728 if (cr->nc.rank_intra == 0)
730 /* Sum with the buffers reversed */
731 MPI_Allreduce(cr->mpb->fbuf, r, nr, MPI_FLOAT, MPI_SUM,
734 MPI_Bcast(r, nr, MPI_FLOAT, 0, cr->nc.comm_intra);
738 MPI_Allreduce(r, cr->mpb->fbuf, nr, MPI_FLOAT, MPI_SUM,
739 cr->mpi_comm_mygroup);
740 for (i = 0; i < nr; i++)
742 r[i] = cr->mpb->fbuf[i];
749 void gmx_sumi(int nr, int r[], const t_commrec *cr)
752 gmx_call("gmx_sumi");
754 #if defined(MPI_IN_PLACE_EXISTS) || defined(GMX_THREAD_MPI)
757 /* Use two step summing */
758 if (cr->nc.rank_intra == 0)
760 MPI_Reduce(MPI_IN_PLACE, r, nr, MPI_INT, MPI_SUM, 0, cr->nc.comm_intra);
761 /* Sum with the buffers reversed */
762 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_INT, MPI_SUM, cr->nc.comm_inter);
766 /* This is here because of the silly MPI specification
767 that MPI_IN_PLACE should be put in sendbuf instead of recvbuf */
768 MPI_Reduce(r, NULL, nr, MPI_INT, MPI_SUM, 0, cr->nc.comm_intra);
770 MPI_Bcast(r, nr, MPI_INT, 0, cr->nc.comm_intra);
774 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_INT, MPI_SUM, cr->mpi_comm_mygroup);
779 if (nr > cr->mpb->ibuf_alloc)
781 cr->mpb->ibuf_alloc = nr;
782 srenew(cr->mpb->ibuf, cr->mpb->ibuf_alloc);
786 /* Use two step summing */
787 MPI_Allreduce(r, cr->mpb->ibuf, nr, MPI_INT, MPI_SUM, cr->nc.comm_intra);
788 if (cr->nc.rank_intra == 0)
790 /* Sum with the buffers reversed */
791 MPI_Allreduce(cr->mpb->ibuf, r, nr, MPI_INT, MPI_SUM, cr->nc.comm_inter);
793 MPI_Bcast(r, nr, MPI_INT, 0, cr->nc.comm_intra);
797 MPI_Allreduce(r, cr->mpb->ibuf, nr, MPI_INT, MPI_SUM, cr->mpi_comm_mygroup);
798 for (i = 0; i < nr; i++)
800 r[i] = cr->mpb->ibuf[i];
807 void gmx_sumli(int nr, gmx_large_int_t r[], const t_commrec *cr)
810 gmx_call("gmx_sumli");
812 #if defined(MPI_IN_PLACE_EXISTS) || defined(GMX_THREAD_MPI)
815 /* Use two step summing */
816 if (cr->nc.rank_intra == 0)
818 MPI_Reduce(MPI_IN_PLACE, r, nr, GMX_MPI_LARGE_INT, MPI_SUM, 0,
820 /* Sum with the buffers reversed */
821 MPI_Allreduce(MPI_IN_PLACE, r, nr, GMX_MPI_LARGE_INT, MPI_SUM,
826 /* This is here because of the silly MPI specification
827 that MPI_IN_PLACE should be put in sendbuf instead of recvbuf */
828 MPI_Reduce(r, NULL, nr, GMX_MPI_LARGE_INT, MPI_SUM, 0, cr->nc.comm_intra);
830 MPI_Bcast(r, nr, GMX_MPI_LARGE_INT, 0, cr->nc.comm_intra);
834 MPI_Allreduce(MPI_IN_PLACE, r, nr, GMX_MPI_LARGE_INT, MPI_SUM, cr->mpi_comm_mygroup);
839 if (nr > cr->mpb->libuf_alloc)
841 cr->mpb->libuf_alloc = nr;
842 srenew(cr->mpb->libuf, cr->mpb->libuf_alloc);
846 /* Use two step summing */
847 MPI_Allreduce(r, cr->mpb->libuf, nr, GMX_MPI_LARGE_INT, MPI_SUM,
849 if (cr->nc.rank_intra == 0)
851 /* Sum with the buffers reversed */
852 MPI_Allreduce(cr->mpb->libuf, r, nr, GMX_MPI_LARGE_INT, MPI_SUM,
855 MPI_Bcast(r, nr, GMX_MPI_LARGE_INT, 0, cr->nc.comm_intra);
859 MPI_Allreduce(r, cr->mpb->libuf, nr, GMX_MPI_LARGE_INT, MPI_SUM,
860 cr->mpi_comm_mygroup);
861 for (i = 0; i < nr; i++)
863 r[i] = cr->mpb->libuf[i];
873 void gmx_sumd_comm(int nr, double r[], MPI_Comm mpi_comm)
875 #if defined(MPI_IN_PLACE_EXISTS) || defined(GMX_THREAD_MPI)
876 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_DOUBLE, MPI_SUM, mpi_comm);
878 /* this function is only used in code that is not performance critical,
879 (during setup, when comm_rec is not the appropriate communication
880 structure), so this isn't as bad as it looks. */
885 MPI_Allreduce(r, buf, nr, MPI_DOUBLE, MPI_SUM, mpi_comm);
886 for (i = 0; i < nr; i++)
896 void gmx_sumf_comm(int nr, float r[], MPI_Comm mpi_comm)
898 #if defined(MPI_IN_PLACE_EXISTS) || defined(GMX_THREAD_MPI)
899 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_FLOAT, MPI_SUM, mpi_comm);
901 /* this function is only used in code that is not performance critical,
902 (during setup, when comm_rec is not the appropriate communication
903 structure), so this isn't as bad as it looks. */
908 MPI_Allreduce(r, buf, nr, MPI_FLOAT, MPI_SUM, mpi_comm);
909 for (i = 0; i < nr; i++)
918 void gmx_sumd_sim(int nr, double r[], const gmx_multisim_t *ms)
921 gmx_call("gmx_sumd_sim");
923 gmx_sumd_comm(nr, r, ms->mpi_comm_masters);
927 void gmx_sumf_sim(int nr, float r[], const gmx_multisim_t *ms)
930 gmx_call("gmx_sumf_sim");
932 gmx_sumf_comm(nr, r, ms->mpi_comm_masters);
936 void gmx_sumi_sim(int nr, int r[], const gmx_multisim_t *ms)
939 gmx_call("gmx_sumi_sim");
941 #if defined(MPI_IN_PLACE_EXISTS) || defined(GMX_THREAD_MPI)
942 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_INT, MPI_SUM, ms->mpi_comm_masters);
944 /* this is thread-unsafe, but it will do for now: */
947 if (nr > ms->mpb->ibuf_alloc)
949 ms->mpb->ibuf_alloc = nr;
950 srenew(ms->mpb->ibuf, ms->mpb->ibuf_alloc);
952 MPI_Allreduce(r, ms->mpb->ibuf, nr, MPI_INT, MPI_SUM, ms->mpi_comm_masters);
953 for (i = 0; i < nr; i++)
955 r[i] = ms->mpb->ibuf[i];
961 void gmx_sumli_sim(int nr, gmx_large_int_t r[], const gmx_multisim_t *ms)
964 gmx_call("gmx_sumli_sim");
966 #if defined(MPI_IN_PLACE_EXISTS) || defined(GMX_THREAD_MPI)
967 MPI_Allreduce(MPI_IN_PLACE, r, nr, GMX_MPI_LARGE_INT, MPI_SUM,
968 ms->mpi_comm_masters);
970 /* this is thread-unsafe, but it will do for now: */
973 if (nr > ms->mpb->libuf_alloc)
975 ms->mpb->libuf_alloc = nr;
976 srenew(ms->mpb->libuf, ms->mpb->libuf_alloc);
978 MPI_Allreduce(r, ms->mpb->libuf, nr, GMX_MPI_LARGE_INT, MPI_SUM,
979 ms->mpi_comm_masters);
980 for (i = 0; i < nr; i++)
982 r[i] = ms->mpb->libuf[i];
989 void gmx_finalize_par(void)
992 /* Compiled without MPI, no MPI finalizing needed */
995 int initialized, finalized;
998 MPI_Initialized(&initialized);
1003 /* just as a check; we don't want to finalize twice */
1004 MPI_Finalized(&finalized);
1010 /* We sync the processes here to try to avoid problems
1011 * with buggy MPI implementations that could cause
1012 * unfinished processes to terminate.
1014 MPI_Barrier(MPI_COMM_WORLD);
1017 if (DOMAINDECOMP(cr)) {
1018 if (cr->npmenodes > 0 || cr->dd->bCartesian)
1019 MPI_Comm_free(&cr->mpi_comm_mygroup);
1020 if (cr->dd->bCartesian)
1021 MPI_Comm_free(&cr->mpi_comm_mysim);
1025 /* Apparently certain mpich implementations cause problems
1026 * with MPI_Finalize. In that case comment out MPI_Finalize.
1030 fprintf(debug, "Will call MPI_Finalize now\n");
1033 ret = MPI_Finalize();
1036 fprintf(debug, "Return code from MPI_Finalize = %d\n", ret);