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 * Copyright (c) 2011,2012,2013,2014,2015,2016,2017,2018, by the GROMACS development team, led by
7 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
8 * and including many others, as listed in the AUTHORS file in the
9 * top-level source directory and at http://www.gromacs.org.
11 * GROMACS is free software; you can redistribute it and/or
12 * modify it under the terms of the GNU Lesser General Public License
13 * as published by the Free Software Foundation; either version 2.1
14 * of the License, or (at your option) any later version.
16 * GROMACS is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
19 * Lesser General Public License for more details.
21 * You should have received a copy of the GNU Lesser General Public
22 * License along with GROMACS; if not, see
23 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
24 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
26 * If you want to redistribute modifications to GROMACS, please
27 * consider that scientific software is very special. Version
28 * control is crucial - bugs must be traceable. We will be happy to
29 * consider code for inclusion in the official distribution, but
30 * derived work must not be called official GROMACS. Details are found
31 * in the README & COPYING files - if they are missing, get the
32 * official version at http://www.gromacs.org.
34 * To help us fund GROMACS development, we humbly ask that you cite
35 * the research papers on the package. Check out http://www.gromacs.org.
48 #include "gromacs/domdec/collect.h"
49 #include "gromacs/gmxlib/network.h"
50 #include "gromacs/math/units.h"
51 #include "gromacs/math/vec.h"
52 #include "gromacs/mdlib/main.h"
53 #include "gromacs/mdtypes/commrec.h"
54 #include "gromacs/mdtypes/enerdata.h"
55 #include "gromacs/mdtypes/inputrec.h"
56 #include "gromacs/mdtypes/md_enums.h"
57 #include "gromacs/mdtypes/state.h"
58 #include "gromacs/random/threefry.h"
59 #include "gromacs/random/uniformintdistribution.h"
60 #include "gromacs/random/uniformrealdistribution.h"
61 #include "gromacs/utility/fatalerror.h"
62 #include "gromacs/utility/pleasecite.h"
63 #include "gromacs/utility/smalloc.h"
66 #define PROBABILITYCUTOFF 100
67 /* we don't bother evaluating if events are more rare than exp(-100) = 3.7x10^-44 */
69 //! Rank in the multisimulaiton
70 #define MSRANK(ms, nodeid) (nodeid)
73 ereTEMP, ereLAMBDA, ereENDSINGLE, ereTL, ereNR
75 static const char *erename[ereNR] = { "temperature", "lambda", "end_single_marker", "temperature and lambda"};
76 /* end_single_marker merely notes the end of single variable replica exchange. All types higher than
77 it are multiple replica exchange methods */
78 /* Eventually, should add 'pressure', 'temperature and pressure', 'lambda_and_pressure', 'temperature_lambda_pressure'?;
79 Let's wait until we feel better about the pressure control methods giving exact ensembles. Right now, we assume constant pressure */
81 typedef struct gmx_repl_ex
83 int repl; /* replica ID */
84 int nrepl; /* total number of replica */
85 real temp; /* temperature */
86 int type; /* replica exchange type from ere enum */
87 real **q; /* quantity, e.g. temperature or lambda; first index is ere, second index is replica ID */
88 gmx_bool bNPT; /* use constant pressure and temperature */
89 real *pres; /* replica pressures */
90 int *ind; /* replica indices */
91 int *allswaps; /* used for keeping track of all the replica swaps */
92 int nst; /* replica exchange interval (number of steps) */
93 int nex; /* number of exchanges per interval */
94 int seed; /* random seed */
95 int nattempt[2]; /* number of even and odd replica change attempts */
96 real *prob_sum; /* sum of probabilities */
97 int **nmoves; /* number of moves between replicas i and j */
98 int *nexchange; /* i-th element of the array is the number of exchanges between replica i-1 and i */
100 /* these are helper arrays for replica exchange; allocated here so they
101 don't have to be allocated each time */
109 /* helper arrays to hold the quantities that are exchanged */
118 static gmx_bool repl_quantity(const gmx_multisim_t *ms,
119 struct gmx_repl_ex *re, int ere, real q)
125 snew(qall, ms->nsim);
127 gmx_sum_sim(ms->nsim, qall, ms);
130 for (s = 1; s < ms->nsim; s++)
132 if (qall[s] != qall[0])
140 /* Set the replica exchange type and quantities */
143 snew(re->q[ere], re->nrepl);
144 for (s = 0; s < ms->nsim; s++)
146 re->q[ere][s] = qall[s];
154 init_replica_exchange(FILE *fplog,
155 const gmx_multisim_t *ms,
156 int numAtomsInSystem,
157 const t_inputrec *ir,
158 const ReplicaExchangeParameters &replExParams)
162 struct gmx_repl_ex *re;
164 gmx_bool bLambda = FALSE;
166 fprintf(fplog, "\nInitializing Replica Exchange\n");
168 if (!isMultiSim(ms) || ms->nsim == 1)
170 gmx_fatal(FARGS, "Nothing to exchange with only one replica, maybe you forgot to set the -multidir option of mdrun?");
172 if (!EI_DYNAMICS(ir->eI))
174 gmx_fatal(FARGS, "Replica exchange is only supported by dynamical simulations");
175 /* Note that PAR(cr) is defined by cr->nnodes > 1, which is
176 * distinct from isMultiSim(ms). A multi-simulation only runs
177 * with real MPI parallelism, but this does not imply PAR(cr)
180 * Since we are using a dynamical integrator, the only
181 * decomposition is DD, so PAR(cr) and DOMAINDECOMP(cr) are
182 * synonymous. The only way for cr->nnodes > 1 to be true is
183 * if we are using DD. */
189 re->nrepl = ms->nsim;
190 snew(re->q, ereENDSINGLE);
192 fprintf(fplog, "Repl There are %d replicas:\n", re->nrepl);
194 /* We only check that the number of atoms in the systms match.
195 * This, of course, do not guarantee that the systems are the same,
196 * but it does guarantee that we can perform replica exchange.
198 check_multi_int(fplog, ms, numAtomsInSystem, "the number of atoms", FALSE);
199 check_multi_int(fplog, ms, ir->eI, "the integrator", FALSE);
200 check_multi_int64(fplog, ms, ir->init_step+ir->nsteps, "init_step+nsteps", FALSE);
201 const int nst = replExParams.exchangeInterval;
202 check_multi_int64(fplog, ms, (ir->init_step+nst-1)/nst,
203 "first exchange step: init_step/-replex", FALSE);
204 check_multi_int(fplog, ms, ir->etc, "the temperature coupling", FALSE);
205 check_multi_int(fplog, ms, ir->opts.ngtc,
206 "the number of temperature coupling groups", FALSE);
207 check_multi_int(fplog, ms, ir->epc, "the pressure coupling", FALSE);
208 check_multi_int(fplog, ms, ir->efep, "free energy", FALSE);
209 check_multi_int(fplog, ms, ir->fepvals->n_lambda, "number of lambda states", FALSE);
211 re->temp = ir->opts.ref_t[0];
212 for (i = 1; (i < ir->opts.ngtc); i++)
214 if (ir->opts.ref_t[i] != re->temp)
216 fprintf(fplog, "\nWARNING: The temperatures of the different temperature coupling groups are not identical\n\n");
217 fprintf(stderr, "\nWARNING: The temperatures of the different temperature coupling groups are not identical\n\n");
222 bTemp = repl_quantity(ms, re, ereTEMP, re->temp);
223 if (ir->efep != efepNO)
225 bLambda = repl_quantity(ms, re, ereLAMBDA, static_cast<real>(ir->fepvals->init_fep_state));
227 if (re->type == -1) /* nothing was assigned */
229 gmx_fatal(FARGS, "The properties of the %d systems are all the same, there is nothing to exchange", re->nrepl);
231 if (bLambda && bTemp)
238 please_cite(fplog, "Sugita1999a");
239 if (ir->epc != epcNO)
242 fprintf(fplog, "Repl Using Constant Pressure REMD.\n");
243 please_cite(fplog, "Okabe2001a");
245 if (ir->etc == etcBERENDSEN)
247 gmx_fatal(FARGS, "REMD with the %s thermostat does not produce correct potential energy distributions, consider using the %s thermostat instead",
248 ETCOUPLTYPE(ir->etc), ETCOUPLTYPE(etcVRESCALE));
253 if (ir->fepvals->delta_lambda != 0) /* check this? */
255 gmx_fatal(FARGS, "delta_lambda is not zero");
260 snew(re->pres, re->nrepl);
261 if (ir->epct == epctSURFACETENSION)
263 pres = ir->ref_p[ZZ][ZZ];
269 for (i = 0; i < DIM; i++)
271 if (ir->compress[i][i] != 0)
273 pres += ir->ref_p[i][i];
279 re->pres[re->repl] = pres;
280 gmx_sum_sim(re->nrepl, re->pres, ms);
283 /* Make an index for increasing replica order */
284 /* only makes sense if one or the other is varying, not both!
285 if both are varying, we trust the order the person gave. */
286 snew(re->ind, re->nrepl);
287 for (i = 0; i < re->nrepl; i++)
292 if (re->type < ereENDSINGLE)
295 for (i = 0; i < re->nrepl; i++)
297 for (j = i+1; j < re->nrepl; j++)
299 if (re->q[re->type][re->ind[j]] < re->q[re->type][re->ind[i]])
301 /* Unordered replicas are supposed to work, but there
302 * is still an issues somewhere.
303 * Note that at this point still re->ind[i]=i.
305 gmx_fatal(FARGS, "Replicas with indices %d < %d have %ss %g > %g, please order your replicas on increasing %s",
308 re->q[re->type][i], re->q[re->type][j],
311 else if (re->q[re->type][re->ind[j]] == re->q[re->type][re->ind[i]])
313 gmx_fatal(FARGS, "Two replicas have identical %ss", erename[re->type]);
319 /* keep track of all the swaps, starting with the initial placement. */
320 snew(re->allswaps, re->nrepl);
321 for (i = 0; i < re->nrepl; i++)
323 re->allswaps[i] = re->ind[i];
329 fprintf(fplog, "\nReplica exchange in temperature\n");
330 for (i = 0; i < re->nrepl; i++)
332 fprintf(fplog, " %5.1f", re->q[re->type][re->ind[i]]);
334 fprintf(fplog, "\n");
337 fprintf(fplog, "\nReplica exchange in lambda\n");
338 for (i = 0; i < re->nrepl; i++)
340 fprintf(fplog, " %3d", static_cast<int>(re->q[re->type][re->ind[i]]));
342 fprintf(fplog, "\n");
345 fprintf(fplog, "\nReplica exchange in temperature and lambda state\n");
346 for (i = 0; i < re->nrepl; i++)
348 fprintf(fplog, " %5.1f", re->q[ereTEMP][re->ind[i]]);
350 fprintf(fplog, "\n");
351 for (i = 0; i < re->nrepl; i++)
353 fprintf(fplog, " %5d", static_cast<int>(re->q[ereLAMBDA][re->ind[i]]));
355 fprintf(fplog, "\n");
358 gmx_incons("Unknown replica exchange quantity");
362 fprintf(fplog, "\nRepl p");
363 for (i = 0; i < re->nrepl; i++)
365 fprintf(fplog, " %5.2f", re->pres[re->ind[i]]);
368 for (i = 0; i < re->nrepl; i++)
370 if ((i > 0) && (re->pres[re->ind[i]] < re->pres[re->ind[i-1]]))
372 fprintf(fplog, "\nWARNING: The reference pressures decrease with increasing temperatures\n\n");
373 fprintf(stderr, "\nWARNING: The reference pressures decrease with increasing temperatures\n\n");
378 if (replExParams.randomSeed == -1)
382 re->seed = static_cast<int>(gmx::makeRandomSeed());
388 gmx_sumi_sim(1, &(re->seed), ms);
392 re->seed = replExParams.randomSeed;
394 fprintf(fplog, "\nReplica exchange interval: %d\n", re->nst);
395 fprintf(fplog, "\nReplica random seed: %d\n", re->seed);
400 snew(re->prob_sum, re->nrepl);
401 snew(re->nexchange, re->nrepl);
402 snew(re->nmoves, re->nrepl);
403 for (i = 0; i < re->nrepl; i++)
405 snew(re->nmoves[i], re->nrepl);
407 fprintf(fplog, "Replica exchange information below: ex and x = exchange, pr = probability\n");
409 /* generate space for the helper functions so we don't have to snew each time */
411 snew(re->destinations, re->nrepl);
412 snew(re->incycle, re->nrepl);
413 snew(re->tmpswap, re->nrepl);
414 snew(re->cyclic, re->nrepl);
415 snew(re->order, re->nrepl);
416 for (i = 0; i < re->nrepl; i++)
418 snew(re->cyclic[i], re->nrepl+1);
419 snew(re->order[i], re->nrepl);
421 /* allocate space for the functions storing the data for the replicas */
422 /* not all of these arrays needed in all cases, but they don't take
423 up much space, since the max size is nrepl**2 */
424 snew(re->prob, re->nrepl);
425 snew(re->bEx, re->nrepl);
426 snew(re->beta, re->nrepl);
427 snew(re->Vol, re->nrepl);
428 snew(re->Epot, re->nrepl);
429 snew(re->de, re->nrepl);
430 for (i = 0; i < re->nrepl; i++)
432 snew(re->de[i], re->nrepl);
434 re->nex = replExParams.numExchanges;
438 static void exchange_reals(const gmx_multisim_t gmx_unused *ms, int gmx_unused b, real *v, int n)
448 MPI_Sendrecv(v, n*sizeof(real),MPI_BYTE,MSRANK(ms,b),0,
449 buf,n*sizeof(real),MPI_BYTE,MSRANK(ms,b),0,
450 ms->mpi_comm_masters,MPI_STATUS_IGNORE);
455 MPI_Isend(v, n*sizeof(real), MPI_BYTE, MSRANK(ms, b), 0,
456 ms->mpi_comm_masters, &mpi_req);
457 MPI_Recv(buf, n*sizeof(real), MPI_BYTE, MSRANK(ms, b), 0,
458 ms->mpi_comm_masters, MPI_STATUS_IGNORE);
459 MPI_Wait(&mpi_req, MPI_STATUS_IGNORE);
462 for (i = 0; i < n; i++)
471 static void exchange_doubles(const gmx_multisim_t gmx_unused *ms, int gmx_unused b, double *v, int n)
481 MPI_Sendrecv(v, n*sizeof(double),MPI_BYTE,MSRANK(ms,b),0,
482 buf,n*sizeof(double),MPI_BYTE,MSRANK(ms,b),0,
483 ms->mpi_comm_masters,MPI_STATUS_IGNORE);
488 MPI_Isend(v, n*sizeof(double), MPI_BYTE, MSRANK(ms, b), 0,
489 ms->mpi_comm_masters, &mpi_req);
490 MPI_Recv(buf, n*sizeof(double), MPI_BYTE, MSRANK(ms, b), 0,
491 ms->mpi_comm_masters, MPI_STATUS_IGNORE);
492 MPI_Wait(&mpi_req, MPI_STATUS_IGNORE);
495 for (i = 0; i < n; i++)
503 static void exchange_rvecs(const gmx_multisim_t gmx_unused *ms, int gmx_unused b, rvec *v, int n)
513 MPI_Sendrecv(v[0], n*sizeof(rvec),MPI_BYTE,MSRANK(ms,b),0,
514 buf[0],n*sizeof(rvec),MPI_BYTE,MSRANK(ms,b),0,
515 ms->mpi_comm_masters,MPI_STATUS_IGNORE);
520 MPI_Isend(v[0], n*sizeof(rvec), MPI_BYTE, MSRANK(ms, b), 0,
521 ms->mpi_comm_masters, &mpi_req);
522 MPI_Recv(buf[0], n*sizeof(rvec), MPI_BYTE, MSRANK(ms, b), 0,
523 ms->mpi_comm_masters, MPI_STATUS_IGNORE);
524 MPI_Wait(&mpi_req, MPI_STATUS_IGNORE);
527 for (i = 0; i < n; i++)
529 copy_rvec(buf[i], v[i]);
535 static void exchange_state(const gmx_multisim_t *ms, int b, t_state *state)
537 /* When t_state changes, this code should be updated. */
539 ngtc = state->ngtc * state->nhchainlength;
540 nnhpres = state->nnhpres* state->nhchainlength;
541 exchange_rvecs(ms, b, state->box, DIM);
542 exchange_rvecs(ms, b, state->box_rel, DIM);
543 exchange_rvecs(ms, b, state->boxv, DIM);
544 exchange_reals(ms, b, &(state->veta), 1);
545 exchange_reals(ms, b, &(state->vol0), 1);
546 exchange_rvecs(ms, b, state->svir_prev, DIM);
547 exchange_rvecs(ms, b, state->fvir_prev, DIM);
548 exchange_rvecs(ms, b, state->pres_prev, DIM);
549 exchange_doubles(ms, b, state->nosehoover_xi.data(), ngtc);
550 exchange_doubles(ms, b, state->nosehoover_vxi.data(), ngtc);
551 exchange_doubles(ms, b, state->nhpres_xi.data(), nnhpres);
552 exchange_doubles(ms, b, state->nhpres_vxi.data(), nnhpres);
553 exchange_doubles(ms, b, state->therm_integral.data(), state->ngtc);
554 exchange_doubles(ms, b, &state->baros_integral, 1);
555 exchange_rvecs(ms, b, as_rvec_array(state->x.data()), state->natoms);
556 exchange_rvecs(ms, b, as_rvec_array(state->v.data()), state->natoms);
559 static void copy_state_serial(const t_state *src, t_state *dest)
563 /* Currently the local state is always a pointer to the global
564 * in serial, so we should never end up here.
565 * TODO: Implement a (trivial) t_state copy once converted to C++.
567 GMX_RELEASE_ASSERT(false, "State copying is currently not implemented in replica exchange");
571 static void scale_velocities(t_state *state, real fac)
575 if (as_rvec_array(state->v.data()))
577 for (i = 0; i < state->natoms; i++)
579 svmul(fac, state->v[i], state->v[i]);
584 static void print_transition_matrix(FILE *fplog, int n, int **nmoves, const int *nattempt)
589 ntot = nattempt[0] + nattempt[1];
590 fprintf(fplog, "\n");
591 fprintf(fplog, "Repl");
592 for (i = 0; i < n; i++)
594 fprintf(fplog, " "); /* put the title closer to the center */
596 fprintf(fplog, "Empirical Transition Matrix\n");
598 fprintf(fplog, "Repl");
599 for (i = 0; i < n; i++)
601 fprintf(fplog, "%8d", (i+1));
603 fprintf(fplog, "\n");
605 for (i = 0; i < n; i++)
607 fprintf(fplog, "Repl");
608 for (j = 0; j < n; j++)
611 if (nmoves[i][j] > 0)
613 Tprint = nmoves[i][j]/(2.0*ntot);
615 fprintf(fplog, "%8.4f", Tprint);
617 fprintf(fplog, "%3d\n", i);
621 static void print_ind(FILE *fplog, const char *leg, int n, int *ind, const gmx_bool *bEx)
625 fprintf(fplog, "Repl %2s %2d", leg, ind[0]);
626 for (i = 1; i < n; i++)
628 fprintf(fplog, " %c %2d", (bEx != nullptr && bEx[i]) ? 'x' : ' ', ind[i]);
630 fprintf(fplog, "\n");
633 static void print_allswitchind(FILE *fplog, int n, int *pind, int *allswaps, int *tmpswap)
637 for (i = 0; i < n; i++)
639 tmpswap[i] = allswaps[i];
641 for (i = 0; i < n; i++)
643 allswaps[i] = tmpswap[pind[i]];
646 fprintf(fplog, "\nAccepted Exchanges: ");
647 for (i = 0; i < n; i++)
649 fprintf(fplog, "%d ", pind[i]);
651 fprintf(fplog, "\n");
653 /* the "Order After Exchange" is the state label corresponding to the configuration that
654 started in state listed in order, i.e.
659 configuration starting in simulation 3 is now in simulation 0,
660 configuration starting in simulation 0 is now in simulation 1,
661 configuration starting in simulation 1 is now in simulation 2,
662 configuration starting in simulation 2 is now in simulation 3
664 fprintf(fplog, "Order After Exchange: ");
665 for (i = 0; i < n; i++)
667 fprintf(fplog, "%d ", allswaps[i]);
669 fprintf(fplog, "\n\n");
672 static void print_prob(FILE *fplog, const char *leg, int n, real *prob)
677 fprintf(fplog, "Repl %2s ", leg);
678 for (i = 1; i < n; i++)
682 sprintf(buf, "%4.2f", prob[i]);
683 fprintf(fplog, " %3s", buf[0] == '1' ? "1.0" : buf+1);
690 fprintf(fplog, "\n");
693 static void print_count(FILE *fplog, const char *leg, int n, int *count)
697 fprintf(fplog, "Repl %2s ", leg);
698 for (i = 1; i < n; i++)
700 fprintf(fplog, " %4d", count[i]);
702 fprintf(fplog, "\n");
705 static real calc_delta(FILE *fplog, gmx_bool bPrint, struct gmx_repl_ex *re, int a, int b, int ap, int bp)
708 real ediff, dpV, delta = 0;
709 real *Epot = re->Epot;
712 real *beta = re->beta;
714 /* Two cases; we are permuted and not. In all cases, setting ap = a and bp = b will reduce
715 to the non permuted case */
721 * Okabe et. al. Chem. Phys. Lett. 335 (2001) 435-439
723 ediff = Epot[b] - Epot[a];
724 delta = -(beta[bp] - beta[ap])*ediff;
727 /* two cases: when we are permuted, and not. */
729 ediff = E_new - E_old
730 = [H_b(x_a) + H_a(x_b)] - [H_b(x_b) + H_a(x_a)]
731 = [H_b(x_a) - H_a(x_a)] + [H_a(x_b) - H_b(x_b)]
732 = de[b][a] + de[a][b] */
735 ediff = E_new - E_old
736 = [H_bp(x_a) + H_ap(x_b)] - [H_bp(x_b) + H_ap(x_a)]
737 = [H_bp(x_a) - H_ap(x_a)] + [H_ap(x_b) - H_bp(x_b)]
738 = [H_bp(x_a) - H_a(x_a) + H_a(x_a) - H_ap(x_a)] + [H_ap(x_b) - H_b(x_b) + H_b(x_b) - H_bp(x_b)]
739 = [H_bp(x_a) - H_a(x_a)] - [H_ap(x_a) - H_a(x_a)] + [H_ap(x_b) - H_b(x_b)] - H_bp(x_b) - H_b(x_b)]
740 = (de[bp][a] - de[ap][a]) + (de[ap][b] - de[bp][b]) */
741 /* but, in the current code implementation, we flip configurations, not indices . . .
742 So let's examine that.
743 = [H_b(x_ap) - H_a(x_a)] - [H_a(x_ap) - H_a(x_a)] + [H_a(x_bp) - H_b(x_b)] - H_b(x_bp) - H_b(x_b)]
744 = [H_b(x_ap) - H_a(x_ap)] + [H_a(x_bp) - H_b(x_pb)]
745 = (de[b][ap] - de[a][ap]) + (de[a][bp] - de[b][bp]
746 So, if we exchange b<=> bp and a<=> ap, we return to the same result.
747 So the simple solution is to flip the
748 position of perturbed and original indices in the tests.
751 ediff = (de[bp][a] - de[ap][a]) + (de[ap][b] - de[bp][b]);
752 delta = ediff*beta[a]; /* assume all same temperature in this case */
756 /* delta = reduced E_new - reduced E_old
757 = [beta_b H_b(x_a) + beta_a H_a(x_b)] - [beta_b H_b(x_b) + beta_a H_a(x_a)]
758 = [beta_b H_b(x_a) - beta_a H_a(x_a)] + [beta_a H_a(x_b) - beta_b H_b(x_b)]
759 = [beta_b dH_b(x_a) + beta_b H_a(x_a) - beta_a H_a(x_a)] +
760 [beta_a dH_a(x_b) + beta_a H_b(x_b) - beta_b H_b(x_b)]
761 = [beta_b dH_b(x_a) + [beta_a dH_a(x_b) +
762 beta_b (H_a(x_a) - H_b(x_b)]) - beta_a (H_a(x_a) - H_b(x_b))
763 = beta_b dH_b(x_a) + beta_a dH_a(x_b) - (beta_b - beta_a)(H_b(x_b) - H_a(x_a) */
764 /* delta = beta[b]*de[b][a] + beta[a]*de[a][b] - (beta[b] - beta[a])*(Epot[b] - Epot[a]; */
765 /* permuted (big breath!) */
766 /* delta = reduced E_new - reduced E_old
767 = [beta_bp H_bp(x_a) + beta_ap H_ap(x_b)] - [beta_bp H_bp(x_b) + beta_ap H_ap(x_a)]
768 = [beta_bp H_bp(x_a) - beta_ap H_ap(x_a)] + [beta_ap H_ap(x_b) - beta_bp H_bp(x_b)]
769 = [beta_bp H_bp(x_a) - beta_ap H_ap(x_a)] + [beta_ap H_ap(x_b) - beta_bp H_bp(x_b)]
770 - beta_pb H_a(x_a) + beta_ap H_a(x_a) + beta_pb H_a(x_a) - beta_ap H_a(x_a)
771 - beta_ap H_b(x_b) + beta_bp H_b(x_b) + beta_ap H_b(x_b) - beta_bp H_b(x_b)
772 = [(beta_bp H_bp(x_a) - beta_bp H_a(x_a)) - (beta_ap H_ap(x_a) - beta_ap H_a(x_a))] +
773 [(beta_ap H_ap(x_b) - beta_ap H_b(x_b)) - (beta_bp H_bp(x_b) - beta_bp H_b(x_b))]
774 + beta_pb H_a(x_a) - beta_ap H_a(x_a) + beta_ap H_b(x_b) - beta_bp H_b(x_b)
775 = [beta_bp (H_bp(x_a) - H_a(x_a)) - beta_ap (H_ap(x_a) - H_a(x_a))] +
776 [beta_ap (H_ap(x_b) - H_b(x_b)) - beta_bp (H_bp(x_b) - H_b(x_b))]
777 + beta_pb (H_a(x_a) - H_b(x_b)) - beta_ap (H_a(x_a) - H_b(x_b))
778 = ([beta_bp de[bp][a] - beta_ap de[ap][a]) + beta_ap de[ap][b] - beta_bp de[bp][b])
779 + (beta_pb-beta_ap)(H_a(x_a) - H_b(x_b)) */
780 delta = beta[bp]*(de[bp][a] - de[bp][b]) + beta[ap]*(de[ap][b] - de[ap][a]) - (beta[bp]-beta[ap])*(Epot[b]-Epot[a]);
783 gmx_incons("Unknown replica exchange quantity");
787 fprintf(fplog, "Repl %d <-> %d dE_term = %10.3e (kT)\n", a, b, delta);
791 /* revist the calculation for 5.0. Might be some improvements. */
792 dpV = (beta[ap]*re->pres[ap]-beta[bp]*re->pres[bp])*(Vol[b]-Vol[a])/PRESFAC;
795 fprintf(fplog, " dpV = %10.3e d = %10.3e\n", dpV, delta + dpV);
803 test_for_replica_exchange(FILE *fplog,
804 const gmx_multisim_t *ms,
805 struct gmx_repl_ex *re,
806 const gmx_enerdata_t *enerd,
811 int m, i, j, a, b, ap, bp, i0, i1, tmp;
813 gmx_bool bPrint, bMultiEx;
814 gmx_bool *bEx = re->bEx;
815 real *prob = re->prob;
816 int *pind = re->destinations; /* permuted index */
817 gmx_bool bEpot = FALSE;
818 gmx_bool bDLambda = FALSE;
819 gmx_bool bVol = FALSE;
820 gmx::ThreeFry2x64<64> rng(re->seed, gmx::RandomDomain::ReplicaExchange);
821 gmx::UniformRealDistribution<real> uniformRealDist;
822 gmx::UniformIntDistribution<int> uniformNreplDist(0, re->nrepl-1);
824 bMultiEx = (re->nex > 1); /* multiple exchanges at each state */
825 fprintf(fplog, "Replica exchange at step %" PRId64 " time %.5f\n", step, time);
829 for (i = 0; i < re->nrepl; i++)
834 re->Vol[re->repl] = vol;
836 if ((re->type == ereTEMP || re->type == ereTL))
838 for (i = 0; i < re->nrepl; i++)
843 re->Epot[re->repl] = enerd->term[F_EPOT];
844 /* temperatures of different states*/
845 for (i = 0; i < re->nrepl; i++)
847 re->beta[i] = 1.0/(re->q[ereTEMP][i]*BOLTZ);
852 for (i = 0; i < re->nrepl; i++)
854 re->beta[i] = 1.0/(re->temp*BOLTZ); /* we have a single temperature */
857 if (re->type == ereLAMBDA || re->type == ereTL)
860 /* lambda differences. */
861 /* de[i][j] is the energy of the jth simulation in the ith Hamiltonian
862 minus the energy of the jth simulation in the jth Hamiltonian */
863 for (i = 0; i < re->nrepl; i++)
865 for (j = 0; j < re->nrepl; j++)
870 for (i = 0; i < re->nrepl; i++)
872 re->de[i][re->repl] = (enerd->enerpart_lambda[static_cast<int>(re->q[ereLAMBDA][i])+1]-enerd->enerpart_lambda[0]);
876 /* now actually do the communication */
879 gmx_sum_sim(re->nrepl, re->Vol, ms);
883 gmx_sum_sim(re->nrepl, re->Epot, ms);
887 for (i = 0; i < re->nrepl; i++)
889 gmx_sum_sim(re->nrepl, re->de[i], ms);
893 /* make a duplicate set of indices for shuffling */
894 for (i = 0; i < re->nrepl; i++)
896 pind[i] = re->ind[i];
899 rng.restart( step, 0 );
903 /* multiple random switch exchange */
907 for (i = 0; i < re->nex + nself; i++)
909 // For now this is superfluous, but just in case we ever add more
910 // calls in different branches it is safer to always reset the distribution.
911 uniformNreplDist.reset();
913 /* randomly select a pair */
914 /* in theory, could reduce this by identifying only which switches had a nonneglibible
915 probability of occurring (log p > -100) and only operate on those switches */
916 /* find out which state it is from, and what label that state currently has. Likely
917 more work that useful. */
918 i0 = uniformNreplDist(rng);
919 i1 = uniformNreplDist(rng);
923 continue; /* self-exchange, back up and do it again */
926 a = re->ind[i0]; /* what are the indices of these states? */
931 bPrint = FALSE; /* too noisy */
932 /* calculate the energy difference */
933 /* if the code changes to flip the STATES, rather than the configurations,
934 use the commented version of the code */
935 /* delta = calc_delta(fplog,bPrint,re,a,b,ap,bp); */
936 delta = calc_delta(fplog, bPrint, re, ap, bp, a, b);
938 /* we actually only use the first space in the prob and bEx array,
939 since there are actually many switches between pairs. */
949 if (delta > PROBABILITYCUTOFF)
955 prob[0] = exp(-delta);
957 // roll a number to determine if accepted. For now it is superfluous to
958 // reset, but just in case we ever add more calls in different branches
959 // it is safer to always reset the distribution.
960 uniformRealDist.reset();
961 bEx[0] = uniformRealDist(rng) < prob[0];
963 re->prob_sum[0] += prob[0];
967 /* swap the states */
973 re->nattempt[0]++; /* keep track of total permutation trials here */
974 print_allswitchind(fplog, re->nrepl, pind, re->allswaps, re->tmpswap);
978 /* standard nearest neighbor replica exchange */
980 m = (step / re->nst) % 2;
981 for (i = 1; i < re->nrepl; i++)
986 bPrint = (re->repl == a || re->repl == b);
989 delta = calc_delta(fplog, bPrint, re, a, b, a, b);
998 if (delta > PROBABILITYCUTOFF)
1004 prob[i] = exp(-delta);
1006 // roll a number to determine if accepted. For now it is superfluous to
1007 // reset, but just in case we ever add more calls in different branches
1008 // it is safer to always reset the distribution.
1009 uniformRealDist.reset();
1010 bEx[i] = uniformRealDist(rng) < prob[i];
1012 re->prob_sum[i] += prob[i];
1016 /* swap these two */
1018 pind[i-1] = pind[i];
1020 re->nexchange[i]++; /* statistics for back compatibility */
1029 /* print some statistics */
1030 print_ind(fplog, "ex", re->nrepl, re->ind, bEx);
1031 print_prob(fplog, "pr", re->nrepl, prob);
1032 fprintf(fplog, "\n");
1036 /* record which moves were made and accepted */
1037 for (i = 0; i < re->nrepl; i++)
1039 re->nmoves[re->ind[i]][pind[i]] += 1;
1040 re->nmoves[pind[i]][re->ind[i]] += 1;
1042 fflush(fplog); /* make sure we can see what the last exchange was */
1046 cyclic_decomposition(const int *destinations,
1055 for (i = 0; i < nrepl; i++)
1059 for (i = 0; i < nrepl; i++) /* one cycle for each replica */
1070 for (j = 0; j < nrepl; j++) /* potentially all cycles are part, but we will break first */
1072 p = destinations[p]; /* start permuting */
1080 break; /* we've reached the original element, the cycle is complete, and we marked the end. */
1084 cyclic[i][c] = p; /* each permutation gives a new member of the cycle */
1090 *nswap = maxlen - 1;
1094 for (i = 0; i < nrepl; i++)
1096 fprintf(debug, "Cycle %d:", i);
1097 for (j = 0; j < nrepl; j++)
1099 if (cyclic[i][j] < 0)
1103 fprintf(debug, "%2d", cyclic[i][j]);
1105 fprintf(debug, "\n");
1112 compute_exchange_order(int **cyclic,
1119 for (j = 0; j < maxswap; j++)
1121 for (i = 0; i < nrepl; i++)
1123 if (cyclic[i][j+1] >= 0)
1125 order[cyclic[i][j+1]][j] = cyclic[i][j];
1126 order[cyclic[i][j]][j] = cyclic[i][j+1];
1129 for (i = 0; i < nrepl; i++)
1131 if (order[i][j] < 0)
1133 order[i][j] = i; /* if it's not exchanging, it should stay this round*/
1140 fprintf(debug, "Replica Exchange Order\n");
1141 for (i = 0; i < nrepl; i++)
1143 fprintf(debug, "Replica %d:", i);
1144 for (j = 0; j < maxswap; j++)
1146 if (order[i][j] < 0)
1150 fprintf(debug, "%2d", order[i][j]);
1152 fprintf(debug, "\n");
1159 prepare_to_do_exchange(struct gmx_repl_ex *re,
1160 const int replica_id,
1162 gmx_bool *bThisReplicaExchanged)
1165 /* Hold the cyclic decomposition of the (multiple) replica
1167 gmx_bool bAnyReplicaExchanged = FALSE;
1168 *bThisReplicaExchanged = FALSE;
1170 for (i = 0; i < re->nrepl; i++)
1172 if (re->destinations[i] != re->ind[i])
1174 /* only mark as exchanged if the index has been shuffled */
1175 bAnyReplicaExchanged = TRUE;
1179 if (bAnyReplicaExchanged)
1181 /* reinitialize the placeholder arrays */
1182 for (i = 0; i < re->nrepl; i++)
1184 for (j = 0; j < re->nrepl; j++)
1186 re->cyclic[i][j] = -1;
1187 re->order[i][j] = -1;
1191 /* Identify the cyclic decomposition of the permutation (very
1192 * fast if neighbor replica exchange). */
1193 cyclic_decomposition(re->destinations, re->cyclic, re->incycle, re->nrepl, maxswap);
1195 /* Now translate the decomposition into a replica exchange
1196 * order at each step. */
1197 compute_exchange_order(re->cyclic, re->order, re->nrepl, *maxswap);
1199 /* Did this replica do any exchange at any point? */
1200 for (j = 0; j < *maxswap; j++)
1202 if (replica_id != re->order[replica_id][j])
1204 *bThisReplicaExchanged = TRUE;
1211 gmx_bool replica_exchange(FILE *fplog, const t_commrec *cr,
1212 const gmx_multisim_t *ms, struct gmx_repl_ex *re,
1213 t_state *state, const gmx_enerdata_t *enerd,
1214 t_state *state_local, int64_t step, real time)
1218 int exchange_partner;
1220 /* Number of rounds of exchanges needed to deal with any multiple
1222 /* Where each replica ends up after the exchange attempt(s). */
1223 /* The order in which multiple exchanges will occur. */
1224 gmx_bool bThisReplicaExchanged = FALSE;
1228 replica_id = re->repl;
1229 test_for_replica_exchange(fplog, ms, re, enerd, det(state_local->box), step, time);
1230 prepare_to_do_exchange(re, replica_id, &maxswap, &bThisReplicaExchanged);
1232 /* Do intra-simulation broadcast so all processors belonging to
1233 * each simulation know whether they need to participate in
1234 * collecting the state. Otherwise, they might as well get on with
1235 * the next thing to do. */
1236 if (DOMAINDECOMP(cr))
1239 MPI_Bcast(&bThisReplicaExchanged, sizeof(gmx_bool), MPI_BYTE, MASTERRANK(cr),
1240 cr->mpi_comm_mygroup);
1244 if (bThisReplicaExchanged)
1246 /* Exchange the states */
1247 /* Collect the global state on the master node */
1248 if (DOMAINDECOMP(cr))
1250 dd_collect_state(cr->dd, state_local, state);
1254 copy_state_serial(state_local, state);
1259 /* There will be only one swap cycle with standard replica
1260 * exchange, but there may be multiple swap cycles if we
1261 * allow multiple swaps. */
1263 for (j = 0; j < maxswap; j++)
1265 exchange_partner = re->order[replica_id][j];
1267 if (exchange_partner != replica_id)
1269 /* Exchange the global states between the master nodes */
1272 fprintf(debug, "Exchanging %d with %d\n", replica_id, exchange_partner);
1274 exchange_state(ms, exchange_partner, state);
1277 /* For temperature-type replica exchange, we need to scale
1278 * the velocities. */
1279 if (re->type == ereTEMP || re->type == ereTL)
1281 scale_velocities(state, std::sqrt(re->q[ereTEMP][replica_id]/re->q[ereTEMP][re->destinations[replica_id]]));
1286 /* With domain decomposition the global state is distributed later */
1287 if (!DOMAINDECOMP(cr))
1289 /* Copy the global state to the local state data structure */
1290 copy_state_serial(state, state_local);
1294 return bThisReplicaExchanged;
1297 void print_replica_exchange_statistics(FILE *fplog, struct gmx_repl_ex *re)
1301 fprintf(fplog, "\nReplica exchange statistics\n");
1305 fprintf(fplog, "Repl %d attempts, %d odd, %d even\n",
1306 re->nattempt[0]+re->nattempt[1], re->nattempt[1], re->nattempt[0]);
1308 fprintf(fplog, "Repl average probabilities:\n");
1309 for (i = 1; i < re->nrepl; i++)
1311 if (re->nattempt[i%2] == 0)
1317 re->prob[i] = re->prob_sum[i]/re->nattempt[i%2];
1320 print_ind(fplog, "", re->nrepl, re->ind, nullptr);
1321 print_prob(fplog, "", re->nrepl, re->prob);
1323 fprintf(fplog, "Repl number of exchanges:\n");
1324 print_ind(fplog, "", re->nrepl, re->ind, nullptr);
1325 print_count(fplog, "", re->nrepl, re->nexchange);
1327 fprintf(fplog, "Repl average number of exchanges:\n");
1328 for (i = 1; i < re->nrepl; i++)
1330 if (re->nattempt[i%2] == 0)
1336 re->prob[i] = (static_cast<real>(re->nexchange[i]))/re->nattempt[i%2];
1339 print_ind(fplog, "", re->nrepl, re->ind, nullptr);
1340 print_prob(fplog, "", re->nrepl, re->prob);
1342 fprintf(fplog, "\n");
1344 /* print the transition matrix */
1345 print_transition_matrix(fplog, re->nrepl, re->nmoves, re->nattempt);