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, 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.
55 #define PROBABILITYCUTOFF 100
56 /* we don't bother evaluating if events are more rare than exp(-100) = 3.7x10^-44 */
59 ereTEMP, ereLAMBDA, ereENDSINGLE, ereTL, ereNR
61 const char *erename[ereNR] = { "temperature", "lambda", "end_single_marker", "temperature and lambda"};
62 /* end_single_marker merely notes the end of single variable replica exchange. All types higher than
63 it are multiple replica exchange methods */
64 /* Eventually, should add 'pressure', 'temperature and pressure', 'lambda_and_pressure', 'temperature_lambda_pressure'?;
65 Let's wait until we feel better about the pressure control methods giving exact ensembles. Right now, we assume constant pressure */
67 typedef struct gmx_repl_ex
86 /* these are helper arrays for replica exchange; allocated here so they
87 don't have to be allocated each time */
95 /* helper arrays to hold the quantities that are exchanged */
104 static gmx_bool repl_quantity(const gmx_multisim_t *ms,
105 struct gmx_repl_ex *re, int ere, real q)
111 snew(qall, ms->nsim);
113 gmx_sum_sim(ms->nsim, qall, ms);
116 for (s = 1; s < ms->nsim; s++)
118 if (qall[s] != qall[0])
126 /* Set the replica exchange type and quantities */
129 snew(re->q[ere], re->nrepl);
130 for (s = 0; s < ms->nsim; s++)
132 re->q[ere][s] = qall[s];
139 gmx_repl_ex_t init_replica_exchange(FILE *fplog,
140 const gmx_multisim_t *ms,
141 const t_state *state,
142 const t_inputrec *ir,
143 int nst, int nex, int init_seed)
147 struct gmx_repl_ex *re;
149 gmx_bool bLambda = FALSE;
151 fprintf(fplog, "\nInitializing Replica Exchange\n");
153 if (ms == NULL || ms->nsim == 1)
155 gmx_fatal(FARGS, "Nothing to exchange with only one replica, maybe you forgot to set the -multi option of mdrun?");
161 re->nrepl = ms->nsim;
162 snew(re->q, ereENDSINGLE);
164 fprintf(fplog, "Repl There are %d replicas:\n", re->nrepl);
166 check_multi_int(fplog, ms, state->natoms, "the number of atoms", FALSE);
167 check_multi_int(fplog, ms, ir->eI, "the integrator", FALSE);
168 check_multi_int64(fplog, ms, ir->init_step+ir->nsteps, "init_step+nsteps", FALSE);
169 check_multi_int64(fplog, ms, (ir->init_step+nst-1)/nst,
170 "first exchange step: init_step/-replex", FALSE);
171 check_multi_int(fplog, ms, ir->etc, "the temperature coupling", FALSE);
172 check_multi_int(fplog, ms, ir->opts.ngtc,
173 "the number of temperature coupling groups", FALSE);
174 check_multi_int(fplog, ms, ir->epc, "the pressure coupling", FALSE);
175 check_multi_int(fplog, ms, ir->efep, "free energy", FALSE);
176 check_multi_int(fplog, ms, ir->fepvals->n_lambda, "number of lambda states", FALSE);
178 re->temp = ir->opts.ref_t[0];
179 for (i = 1; (i < ir->opts.ngtc); i++)
181 if (ir->opts.ref_t[i] != re->temp)
183 fprintf(fplog, "\nWARNING: The temperatures of the different temperature coupling groups are not identical\n\n");
184 fprintf(stderr, "\nWARNING: The temperatures of the different temperature coupling groups are not identical\n\n");
189 bTemp = repl_quantity(ms, re, ereTEMP, re->temp);
190 if (ir->efep != efepNO)
192 bLambda = repl_quantity(ms, re, ereLAMBDA, (real)ir->fepvals->init_fep_state);
194 if (re->type == -1) /* nothing was assigned */
196 gmx_fatal(FARGS, "The properties of the %d systems are all the same, there is nothing to exchange", re->nrepl);
198 if (bLambda && bTemp)
205 please_cite(fplog, "Sugita1999a");
206 if (ir->epc != epcNO)
209 fprintf(fplog, "Repl Using Constant Pressure REMD.\n");
210 please_cite(fplog, "Okabe2001a");
212 if (ir->etc == etcBERENDSEN)
214 gmx_fatal(FARGS, "REMD with the %s thermostat does not produce correct potential energy distributions, consider using the %s thermostat instead",
215 ETCOUPLTYPE(ir->etc), ETCOUPLTYPE(etcVRESCALE));
220 if (ir->fepvals->delta_lambda != 0) /* check this? */
222 gmx_fatal(FARGS, "delta_lambda is not zero");
227 snew(re->pres, re->nrepl);
228 if (ir->epct == epctSURFACETENSION)
230 pres = ir->ref_p[ZZ][ZZ];
236 for (i = 0; i < DIM; i++)
238 if (ir->compress[i][i] != 0)
240 pres += ir->ref_p[i][i];
246 re->pres[re->repl] = pres;
247 gmx_sum_sim(re->nrepl, re->pres, ms);
250 /* Make an index for increasing replica order */
251 /* only makes sense if one or the other is varying, not both!
252 if both are varying, we trust the order the person gave. */
253 snew(re->ind, re->nrepl);
254 for (i = 0; i < re->nrepl; i++)
259 if (re->type < ereENDSINGLE)
262 for (i = 0; i < re->nrepl; i++)
264 for (j = i+1; j < re->nrepl; j++)
266 if (re->q[re->type][re->ind[j]] < re->q[re->type][re->ind[i]])
269 re->ind[i] = re->ind[j];
272 else if (re->q[re->type][re->ind[j]] == re->q[re->type][re->ind[i]])
274 gmx_fatal(FARGS, "Two replicas have identical %ss", erename[re->type]);
280 /* keep track of all the swaps, starting with the initial placement. */
281 snew(re->allswaps, re->nrepl);
282 for (i = 0; i < re->nrepl; i++)
284 re->allswaps[i] = re->ind[i];
290 fprintf(fplog, "\nReplica exchange in temperature\n");
291 for (i = 0; i < re->nrepl; i++)
293 fprintf(fplog, " %5.1f", re->q[re->type][re->ind[i]]);
295 fprintf(fplog, "\n");
298 fprintf(fplog, "\nReplica exchange in lambda\n");
299 for (i = 0; i < re->nrepl; i++)
301 fprintf(fplog, " %3d", (int)re->q[re->type][re->ind[i]]);
303 fprintf(fplog, "\n");
306 fprintf(fplog, "\nReplica exchange in temperature and lambda state\n");
307 for (i = 0; i < re->nrepl; i++)
309 fprintf(fplog, " %5.1f", re->q[ereTEMP][re->ind[i]]);
311 fprintf(fplog, "\n");
312 for (i = 0; i < re->nrepl; i++)
314 fprintf(fplog, " %5d", (int)re->q[ereLAMBDA][re->ind[i]]);
316 fprintf(fplog, "\n");
319 gmx_incons("Unknown replica exchange quantity");
323 fprintf(fplog, "\nRepl p");
324 for (i = 0; i < re->nrepl; i++)
326 fprintf(fplog, " %5.2f", re->pres[re->ind[i]]);
329 for (i = 0; i < re->nrepl; i++)
331 if ((i > 0) && (re->pres[re->ind[i]] < re->pres[re->ind[i-1]]))
333 fprintf(fplog, "\nWARNING: The reference pressures decrease with increasing temperatures\n\n");
334 fprintf(stderr, "\nWARNING: The reference pressures decrease with increasing temperatures\n\n");
343 re->seed = make_seed();
349 gmx_sumi_sim(1, &(re->seed), ms);
353 re->seed = init_seed;
355 fprintf(fplog, "\nReplica exchange interval: %d\n", re->nst);
356 fprintf(fplog, "\nReplica random seed: %d\n", re->seed);
361 snew(re->prob_sum, re->nrepl);
362 snew(re->nexchange, re->nrepl);
363 snew(re->nmoves, re->nrepl);
364 for (i = 0; i < re->nrepl; i++)
366 snew(re->nmoves[i], re->nrepl);
368 fprintf(fplog, "Replica exchange information below: x=exchange, pr=probability\n");
370 /* generate space for the helper functions so we don't have to snew each time */
372 snew(re->destinations, re->nrepl);
373 snew(re->incycle, re->nrepl);
374 snew(re->tmpswap, re->nrepl);
375 snew(re->cyclic, re->nrepl);
376 snew(re->order, re->nrepl);
377 for (i = 0; i < re->nrepl; i++)
379 snew(re->cyclic[i], re->nrepl);
380 snew(re->order[i], re->nrepl);
382 /* allocate space for the functions storing the data for the replicas */
383 /* not all of these arrays needed in all cases, but they don't take
384 up much space, since the max size is nrepl**2 */
385 snew(re->prob, re->nrepl);
386 snew(re->bEx, re->nrepl);
387 snew(re->beta, re->nrepl);
388 snew(re->Vol, re->nrepl);
389 snew(re->Epot, re->nrepl);
390 snew(re->de, re->nrepl);
391 for (i = 0; i < re->nrepl; i++)
393 snew(re->de[i], re->nrepl);
399 static void exchange_reals(const gmx_multisim_t gmx_unused *ms, int gmx_unused b, real *v, int n)
409 MPI_Sendrecv(v, n*sizeof(real),MPI_BYTE,MSRANK(ms,b),0,
410 buf,n*sizeof(real),MPI_BYTE,MSRANK(ms,b),0,
411 ms->mpi_comm_masters,MPI_STATUS_IGNORE);
416 MPI_Isend(v, n*sizeof(real), MPI_BYTE, MSRANK(ms, b), 0,
417 ms->mpi_comm_masters, &mpi_req);
418 MPI_Recv(buf, n*sizeof(real), MPI_BYTE, MSRANK(ms, b), 0,
419 ms->mpi_comm_masters, MPI_STATUS_IGNORE);
420 MPI_Wait(&mpi_req, MPI_STATUS_IGNORE);
423 for (i = 0; i < n; i++)
432 static void exchange_ints(const gmx_multisim_t gmx_unused *ms, int gmx_unused b, int *v, int n)
442 MPI_Sendrecv(v, n*sizeof(int),MPI_BYTE,MSRANK(ms,b),0,
443 buf,n*sizeof(int),MPI_BYTE,MSRANK(ms,b),0,
444 ms->mpi_comm_masters,MPI_STATUS_IGNORE);
449 MPI_Isend(v, n*sizeof(int), MPI_BYTE, MSRANK(ms, b), 0,
450 ms->mpi_comm_masters, &mpi_req);
451 MPI_Recv(buf, n*sizeof(int), MPI_BYTE, MSRANK(ms, b), 0,
452 ms->mpi_comm_masters, MPI_STATUS_IGNORE);
453 MPI_Wait(&mpi_req, MPI_STATUS_IGNORE);
456 for (i = 0; i < n; i++)
464 static void exchange_doubles(const gmx_multisim_t gmx_unused *ms, int gmx_unused b, double *v, int n)
474 MPI_Sendrecv(v, n*sizeof(double),MPI_BYTE,MSRANK(ms,b),0,
475 buf,n*sizeof(double),MPI_BYTE,MSRANK(ms,b),0,
476 ms->mpi_comm_masters,MPI_STATUS_IGNORE);
481 MPI_Isend(v, n*sizeof(double), MPI_BYTE, MSRANK(ms, b), 0,
482 ms->mpi_comm_masters, &mpi_req);
483 MPI_Recv(buf, n*sizeof(double), MPI_BYTE, MSRANK(ms, b), 0,
484 ms->mpi_comm_masters, MPI_STATUS_IGNORE);
485 MPI_Wait(&mpi_req, MPI_STATUS_IGNORE);
488 for (i = 0; i < n; i++)
496 static void exchange_rvecs(const gmx_multisim_t gmx_unused *ms, int gmx_unused b, rvec *v, int n)
506 MPI_Sendrecv(v[0], n*sizeof(rvec),MPI_BYTE,MSRANK(ms,b),0,
507 buf[0],n*sizeof(rvec),MPI_BYTE,MSRANK(ms,b),0,
508 ms->mpi_comm_masters,MPI_STATUS_IGNORE);
513 MPI_Isend(v[0], n*sizeof(rvec), MPI_BYTE, MSRANK(ms, b), 0,
514 ms->mpi_comm_masters, &mpi_req);
515 MPI_Recv(buf[0], n*sizeof(rvec), MPI_BYTE, MSRANK(ms, b), 0,
516 ms->mpi_comm_masters, MPI_STATUS_IGNORE);
517 MPI_Wait(&mpi_req, MPI_STATUS_IGNORE);
520 for (i = 0; i < n; i++)
522 copy_rvec(buf[i], v[i]);
528 static void exchange_state(const gmx_multisim_t *ms, int b, t_state *state)
530 /* When t_state changes, this code should be updated. */
532 ngtc = state->ngtc * state->nhchainlength;
533 nnhpres = state->nnhpres* state->nhchainlength;
534 exchange_rvecs(ms, b, state->box, DIM);
535 exchange_rvecs(ms, b, state->box_rel, DIM);
536 exchange_rvecs(ms, b, state->boxv, DIM);
537 exchange_reals(ms, b, &(state->veta), 1);
538 exchange_reals(ms, b, &(state->vol0), 1);
539 exchange_rvecs(ms, b, state->svir_prev, DIM);
540 exchange_rvecs(ms, b, state->fvir_prev, DIM);
541 exchange_rvecs(ms, b, state->pres_prev, DIM);
542 exchange_doubles(ms, b, state->nosehoover_xi, ngtc);
543 exchange_doubles(ms, b, state->nosehoover_vxi, ngtc);
544 exchange_doubles(ms, b, state->nhpres_xi, nnhpres);
545 exchange_doubles(ms, b, state->nhpres_vxi, nnhpres);
546 exchange_doubles(ms, b, state->therm_integral, state->ngtc);
547 exchange_rvecs(ms, b, state->x, state->natoms);
548 exchange_rvecs(ms, b, state->v, state->natoms);
549 exchange_rvecs(ms, b, state->sd_X, state->natoms);
552 static void copy_rvecs(rvec *s, rvec *d, int n)
558 for (i = 0; i < n; i++)
560 copy_rvec(s[i], d[i]);
565 static void copy_doubles(const double *s, double *d, int n)
571 for (i = 0; i < n; i++)
578 static void copy_reals(const real *s, real *d, int n)
584 for (i = 0; i < n; i++)
591 static void copy_ints(const int *s, int *d, int n)
597 for (i = 0; i < n; i++)
604 #define scopy_rvecs(v, n) copy_rvecs(state->v, state_local->v, n);
605 #define scopy_doubles(v, n) copy_doubles(state->v, state_local->v, n);
606 #define scopy_reals(v, n) copy_reals(state->v, state_local->v, n);
607 #define scopy_ints(v, n) copy_ints(state->v, state_local->v, n);
609 static void copy_state_nonatomdata(t_state *state, t_state *state_local)
611 /* When t_state changes, this code should be updated. */
613 ngtc = state->ngtc * state->nhchainlength;
614 nnhpres = state->nnhpres* state->nhchainlength;
615 scopy_rvecs(box, DIM);
616 scopy_rvecs(box_rel, DIM);
617 scopy_rvecs(boxv, DIM);
618 state_local->veta = state->veta;
619 state_local->vol0 = state->vol0;
620 scopy_rvecs(svir_prev, DIM);
621 scopy_rvecs(fvir_prev, DIM);
622 scopy_rvecs(pres_prev, DIM);
623 scopy_doubles(nosehoover_xi, ngtc);
624 scopy_doubles(nosehoover_vxi, ngtc);
625 scopy_doubles(nhpres_xi, nnhpres);
626 scopy_doubles(nhpres_vxi, nnhpres);
627 scopy_doubles(therm_integral, state->ngtc);
628 scopy_rvecs(x, state->natoms);
629 scopy_rvecs(v, state->natoms);
630 scopy_rvecs(sd_X, state->natoms);
631 copy_ints(&(state->fep_state), &(state_local->fep_state), 1);
632 scopy_reals(lambda, efptNR);
635 static void scale_velocities(t_state *state, real fac)
641 for (i = 0; i < state->natoms; i++)
643 svmul(fac, state->v[i], state->v[i]);
648 static void pd_collect_state(const t_commrec *cr, t_state *state)
654 fprintf(debug, "Collecting state before exchange\n");
656 shift = cr->nnodes - cr->npmenodes - 1;
657 move_rvecs(cr, FALSE, FALSE, state->x, NULL, shift, NULL);
660 move_rvecs(cr, FALSE, FALSE, state->v, NULL, shift, NULL);
664 move_rvecs(cr, FALSE, FALSE, state->sd_X, NULL, shift, NULL);
668 static void print_transition_matrix(FILE *fplog, int n, int **nmoves, int *nattempt)
673 ntot = nattempt[0] + nattempt[1];
674 fprintf(fplog, "\n");
675 fprintf(fplog, "Repl");
676 for (i = 0; i < n; i++)
678 fprintf(fplog, " "); /* put the title closer to the center */
680 fprintf(fplog, "Empirical Transition Matrix\n");
682 fprintf(fplog, "Repl");
683 for (i = 0; i < n; i++)
685 fprintf(fplog, "%8d", (i+1));
687 fprintf(fplog, "\n");
689 for (i = 0; i < n; i++)
691 fprintf(fplog, "Repl");
692 for (j = 0; j < n; j++)
695 if (nmoves[i][j] > 0)
697 Tprint = nmoves[i][j]/(2.0*ntot);
699 fprintf(fplog, "%8.4f", Tprint);
701 fprintf(fplog, "%3d\n", i);
705 static void print_ind(FILE *fplog, const char *leg, int n, int *ind, gmx_bool *bEx)
709 fprintf(fplog, "Repl %2s %2d", leg, ind[0]);
710 for (i = 1; i < n; i++)
712 fprintf(fplog, " %c %2d", (bEx != 0 && bEx[i]) ? 'x' : ' ', ind[i]);
714 fprintf(fplog, "\n");
717 static void print_allswitchind(FILE *fplog, int n, int *pind, int *allswaps, int *tmpswap)
721 for (i = 0; i < n; i++)
723 tmpswap[i] = allswaps[i];
725 for (i = 0; i < n; i++)
727 allswaps[i] = tmpswap[pind[i]];
730 fprintf(fplog, "\nAccepted Exchanges: ");
731 for (i = 0; i < n; i++)
733 fprintf(fplog, "%d ", pind[i]);
735 fprintf(fplog, "\n");
737 /* the "Order After Exchange" is the state label corresponding to the configuration that
738 started in state listed in order, i.e.
743 configuration starting in simulation 3 is now in simulation 0,
744 configuration starting in simulation 0 is now in simulation 1,
745 configuration starting in simulation 1 is now in simulation 2,
746 configuration starting in simulation 2 is now in simulation 3
748 fprintf(fplog, "Order After Exchange: ");
749 for (i = 0; i < n; i++)
751 fprintf(fplog, "%d ", allswaps[i]);
753 fprintf(fplog, "\n\n");
756 static void print_prob(FILE *fplog, const char *leg, int n, real *prob)
761 fprintf(fplog, "Repl %2s ", leg);
762 for (i = 1; i < n; i++)
766 sprintf(buf, "%4.2f", prob[i]);
767 fprintf(fplog, " %3s", buf[0] == '1' ? "1.0" : buf+1);
774 fprintf(fplog, "\n");
777 static void print_count(FILE *fplog, const char *leg, int n, int *count)
781 fprintf(fplog, "Repl %2s ", leg);
782 for (i = 1; i < n; i++)
784 fprintf(fplog, " %4d", count[i]);
786 fprintf(fplog, "\n");
789 static real calc_delta(FILE *fplog, gmx_bool bPrint, struct gmx_repl_ex *re, int a, int b, int ap, int bp)
792 real ediff, dpV, delta = 0;
793 real *Epot = re->Epot;
796 real *beta = re->beta;
798 /* Two cases; we are permuted and not. In all cases, setting ap = a and bp = b will reduce
799 to the non permuted case */
805 * Okabe et. al. Chem. Phys. Lett. 335 (2001) 435-439
807 ediff = Epot[b] - Epot[a];
808 delta = -(beta[bp] - beta[ap])*ediff;
811 /* two cases: when we are permuted, and not. */
813 ediff = E_new - E_old
814 = [H_b(x_a) + H_a(x_b)] - [H_b(x_b) + H_a(x_a)]
815 = [H_b(x_a) - H_a(x_a)] + [H_a(x_b) - H_b(x_b)]
816 = de[b][a] + de[a][b] */
819 ediff = E_new - E_old
820 = [H_bp(x_a) + H_ap(x_b)] - [H_bp(x_b) + H_ap(x_a)]
821 = [H_bp(x_a) - H_ap(x_a)] + [H_ap(x_b) - H_bp(x_b)]
822 = [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)]
823 = [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)]
824 = (de[bp][a] - de[ap][a]) + (de[ap][b] - de[bp][b]) */
825 /* but, in the current code implementation, we flip configurations, not indices . . .
826 So let's examine that.
827 = [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)]
828 = [H_b(x_ap) - H_a(x_ap)] + [H_a(x_bp) - H_b(x_pb)]
829 = (de[b][ap] - de[a][ap]) + (de[a][bp] - de[b][bp]
830 So, if we exchange b<=> bp and a<=> ap, we return to the same result.
831 So the simple solution is to flip the
832 position of perturbed and original indices in the tests.
835 ediff = (de[bp][a] - de[ap][a]) + (de[ap][b] - de[bp][b]);
836 delta = ediff*beta[a]; /* assume all same temperature in this case */
840 /* delta = reduced E_new - reduced E_old
841 = [beta_b H_b(x_a) + beta_a H_a(x_b)] - [beta_b H_b(x_b) + beta_a H_a(x_a)]
842 = [beta_b H_b(x_a) - beta_a H_a(x_a)] + [beta_a H_a(x_b) - beta_b H_b(x_b)]
843 = [beta_b dH_b(x_a) + beta_b H_a(x_a) - beta_a H_a(x_a)] +
844 [beta_a dH_a(x_b) + beta_a H_b(x_b) - beta_b H_b(x_b)]
845 = [beta_b dH_b(x_a) + [beta_a dH_a(x_b) +
846 beta_b (H_a(x_a) - H_b(x_b)]) - beta_a (H_a(x_a) - H_b(x_b))
847 = beta_b dH_b(x_a) + beta_a dH_a(x_b) - (beta_b - beta_a)(H_b(x_b) - H_a(x_a) */
848 /* delta = beta[b]*de[b][a] + beta[a]*de[a][b] - (beta[b] - beta[a])*(Epot[b] - Epot[a]; */
849 /* permuted (big breath!) */
850 /* delta = reduced E_new - reduced E_old
851 = [beta_bp H_bp(x_a) + beta_ap H_ap(x_b)] - [beta_bp H_bp(x_b) + beta_ap H_ap(x_a)]
852 = [beta_bp H_bp(x_a) - beta_ap H_ap(x_a)] + [beta_ap H_ap(x_b) - beta_bp H_bp(x_b)]
853 = [beta_bp H_bp(x_a) - beta_ap H_ap(x_a)] + [beta_ap H_ap(x_b) - beta_bp H_bp(x_b)]
854 - beta_pb H_a(x_a) + beta_ap H_a(x_a) + beta_pb H_a(x_a) - beta_ap H_a(x_a)
855 - beta_ap H_b(x_b) + beta_bp H_b(x_b) + beta_ap H_b(x_b) - beta_bp H_b(x_b)
856 = [(beta_bp H_bp(x_a) - beta_bp H_a(x_a)) - (beta_ap H_ap(x_a) - beta_ap H_a(x_a))] +
857 [(beta_ap H_ap(x_b) - beta_ap H_b(x_b)) - (beta_bp H_bp(x_b) - beta_bp H_b(x_b))]
858 + beta_pb H_a(x_a) - beta_ap H_a(x_a) + beta_ap H_b(x_b) - beta_bp H_b(x_b)
859 = [beta_bp (H_bp(x_a) - H_a(x_a)) - beta_ap (H_ap(x_a) - H_a(x_a))] +
860 [beta_ap (H_ap(x_b) - H_b(x_b)) - beta_bp (H_bp(x_b) - H_b(x_b))]
861 + beta_pb (H_a(x_a) - H_b(x_b)) - beta_ap (H_a(x_a) - H_b(x_b))
862 = ([beta_bp de[bp][a] - beta_ap de[ap][a]) + beta_ap de[ap][b] - beta_bp de[bp][b])
863 + (beta_pb-beta_ap)(H_a(x_a) - H_b(x_b)) */
864 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]);
867 gmx_incons("Unknown replica exchange quantity");
871 fprintf(fplog, "Repl %d <-> %d dE_term = %10.3e (kT)\n", a, b, delta);
875 /* revist the calculation for 5.0. Might be some improvements. */
876 dpV = (beta[ap]*re->pres[ap]-beta[bp]*re->pres[bp])*(Vol[b]-Vol[a])/PRESFAC;
879 fprintf(fplog, " dpV = %10.3e d = %10.3e\nb", dpV, delta + dpV);
887 test_for_replica_exchange(FILE *fplog,
888 const gmx_multisim_t *ms,
889 struct gmx_repl_ex *re,
890 gmx_enerdata_t *enerd,
895 int m, i, j, a, b, ap, bp, i0, i1, tmp;
896 real ediff = 0, delta = 0, dpV = 0;
897 gmx_bool bPrint, bMultiEx;
898 gmx_bool *bEx = re->bEx;
899 real *prob = re->prob;
900 int *pind = re->destinations; /* permuted index */
901 gmx_bool bEpot = FALSE;
902 gmx_bool bDLambda = FALSE;
903 gmx_bool bVol = FALSE;
905 bMultiEx = (re->nex > 1); /* multiple exchanges at each state */
906 fprintf(fplog, "Replica exchange at step " "%"GMX_PRId64 " time %g\n", step, time);
910 for (i = 0; i < re->nrepl; i++)
915 re->Vol[re->repl] = vol;
917 if ((re->type == ereTEMP || re->type == ereTL))
919 for (i = 0; i < re->nrepl; i++)
924 re->Epot[re->repl] = enerd->term[F_EPOT];
925 /* temperatures of different states*/
926 for (i = 0; i < re->nrepl; i++)
928 re->beta[i] = 1.0/(re->q[ereTEMP][i]*BOLTZ);
933 for (i = 0; i < re->nrepl; i++)
935 re->beta[i] = 1.0/(re->temp*BOLTZ); /* we have a single temperature */
938 if (re->type == ereLAMBDA || re->type == ereTL)
941 /* lambda differences. */
942 /* de[i][j] is the energy of the jth simulation in the ith Hamiltonian
943 minus the energy of the jth simulation in the jth Hamiltonian */
944 for (i = 0; i < re->nrepl; i++)
946 for (j = 0; j < re->nrepl; j++)
951 for (i = 0; i < re->nrepl; i++)
953 re->de[i][re->repl] = (enerd->enerpart_lambda[(int)re->q[ereLAMBDA][i]+1]-enerd->enerpart_lambda[0]);
957 /* now actually do the communication */
960 gmx_sum_sim(re->nrepl, re->Vol, ms);
964 gmx_sum_sim(re->nrepl, re->Epot, ms);
968 for (i = 0; i < re->nrepl; i++)
970 gmx_sum_sim(re->nrepl, re->de[i], ms);
974 /* make a duplicate set of indices for shuffling */
975 for (i = 0; i < re->nrepl; i++)
977 pind[i] = re->ind[i];
982 /* multiple random switch exchange */
983 for (i = 0; i < re->nex; i++)
985 /* randomly select a pair */
986 /* in theory, could reduce this by identifying only which switches had a nonneglibible
987 probability of occurring (log p > -100) and only operate on those switches */
988 /* find out which state it is from, and what label that state currently has. Likely
989 more work that useful. */
990 i0 = (int)(re->nrepl*rando(&(re->seed)));
991 i1 = (int)(re->nrepl*rando(&(re->seed)));
995 continue; /* self-exchange, back up and do it again */
998 a = re->ind[i0]; /* what are the indices of these states? */
1003 bPrint = FALSE; /* too noisy */
1004 /* calculate the energy difference */
1005 /* if the code changes to flip the STATES, rather than the configurations,
1006 use the commented version of the code */
1007 /* delta = calc_delta(fplog,bPrint,re,a,b,ap,bp); */
1008 delta = calc_delta(fplog, bPrint, re, ap, bp, a, b);
1010 /* we actually only use the first space in the prob and bEx array,
1011 since there are actually many switches between pairs. */
1021 if (delta > PROBABILITYCUTOFF)
1027 prob[0] = exp(-delta);
1029 /* roll a number to determine if accepted */
1030 bEx[0] = (rando(&(re->seed)) < prob[0]);
1032 re->prob_sum[0] += prob[0];
1036 /* swap the states */
1038 pind[i0] = pind[i1];
1042 re->nattempt[0]++; /* keep track of total permutation trials here */
1043 print_allswitchind(fplog, re->nrepl, pind, re->allswaps, re->tmpswap);
1047 /* standard nearest neighbor replica exchange */
1048 m = (step / re->nst) % 2;
1049 for (i = 1; i < re->nrepl; i++)
1054 bPrint = (re->repl == a || re->repl == b);
1057 delta = calc_delta(fplog, bPrint, re, a, b, a, b);
1066 if (delta > PROBABILITYCUTOFF)
1072 prob[i] = exp(-delta);
1074 /* roll a number to determine if accepted */
1075 bEx[i] = (rando(&(re->seed)) < prob[i]);
1077 re->prob_sum[i] += prob[i];
1081 /* swap these two */
1083 pind[i-1] = pind[i];
1085 re->nexchange[i]++; /* statistics for back compatibility */
1094 /* print some statistics */
1095 print_ind(fplog, "ex", re->nrepl, re->ind, bEx);
1096 print_prob(fplog, "pr", re->nrepl, prob);
1097 fprintf(fplog, "\n");
1101 /* record which moves were made and accepted */
1102 for (i = 0; i < re->nrepl; i++)
1104 re->nmoves[re->ind[i]][pind[i]] += 1;
1105 re->nmoves[pind[i]][re->ind[i]] += 1;
1107 fflush(fplog); /* make sure we can see what the last exchange was */
1110 static void write_debug_x(t_state *state)
1116 for (i = 0; i < state->natoms; i += 10)
1118 fprintf(debug, "dx %5d %10.5f %10.5f %10.5f\n", i, state->x[i][XX], state->x[i][YY], state->x[i][ZZ]);
1124 cyclic_decomposition(const int *destinations,
1133 for (i = 0; i < nrepl; i++)
1137 for (i = 0; i < nrepl; i++) /* one cycle for each replica */
1148 for (j = 0; j < nrepl; j++) /* potentially all cycles are part, but we will break first */
1150 p = destinations[p]; /* start permuting */
1158 break; /* we've reached the original element, the cycle is complete, and we marked the end. */
1162 cyclic[i][c] = p; /* each permutation gives a new member of the cycle */
1168 *nswap = maxlen - 1;
1172 for (i = 0; i < nrepl; i++)
1174 fprintf(debug, "Cycle %d:", i);
1175 for (j = 0; j < nrepl; j++)
1177 if (cyclic[i][j] < 0)
1181 fprintf(debug, "%2d", cyclic[i][j]);
1183 fprintf(debug, "\n");
1190 compute_exchange_order(FILE *fplog,
1198 for (j = 0; j < maxswap; j++)
1200 for (i = 0; i < nrepl; i++)
1202 if (cyclic[i][j+1] >= 0)
1204 order[cyclic[i][j+1]][j] = cyclic[i][j];
1205 order[cyclic[i][j]][j] = cyclic[i][j+1];
1208 for (i = 0; i < nrepl; i++)
1210 if (order[i][j] < 0)
1212 order[i][j] = i; /* if it's not exchanging, it should stay this round*/
1219 fprintf(fplog, "Replica Exchange Order\n");
1220 for (i = 0; i < nrepl; i++)
1222 fprintf(fplog, "Replica %d:", i);
1223 for (j = 0; j < maxswap; j++)
1225 if (order[i][j] < 0)
1229 fprintf(debug, "%2d", order[i][j]);
1231 fprintf(fplog, "\n");
1238 prepare_to_do_exchange(FILE *fplog,
1239 const int *destinations,
1240 const int replica_id,
1246 gmx_bool *bThisReplicaExchanged)
1249 /* Hold the cyclic decomposition of the (multiple) replica
1251 gmx_bool bAnyReplicaExchanged = FALSE;
1252 *bThisReplicaExchanged = FALSE;
1254 for (i = 0; i < nrepl; i++)
1256 if (destinations[i] != i)
1258 /* only mark as exchanged if the index has been shuffled */
1259 bAnyReplicaExchanged = TRUE;
1263 if (bAnyReplicaExchanged)
1265 /* reinitialize the placeholder arrays */
1266 for (i = 0; i < nrepl; i++)
1268 for (j = 0; j < nrepl; j++)
1275 /* Identify the cyclic decomposition of the permutation (very
1276 * fast if neighbor replica exchange). */
1277 cyclic_decomposition(destinations, cyclic, incycle, nrepl, maxswap);
1279 /* Now translate the decomposition into a replica exchange
1280 * order at each step. */
1281 compute_exchange_order(fplog, cyclic, order, nrepl, *maxswap);
1283 /* Did this replica do any exchange at any point? */
1284 for (j = 0; j < *maxswap; j++)
1286 if (replica_id != order[replica_id][j])
1288 *bThisReplicaExchanged = TRUE;
1295 gmx_bool replica_exchange(FILE *fplog, const t_commrec *cr, struct gmx_repl_ex *re,
1296 t_state *state, gmx_enerdata_t *enerd,
1297 t_state *state_local, gmx_int64_t step, real time)
1301 int exchange_partner;
1303 /* Number of rounds of exchanges needed to deal with any multiple
1305 /* Where each replica ends up after the exchange attempt(s). */
1306 /* The order in which multiple exchanges will occur. */
1307 gmx_bool bThisReplicaExchanged = FALSE;
1311 replica_id = re->repl;
1312 test_for_replica_exchange(fplog, cr->ms, re, enerd, det(state_local->box), step, time);
1313 prepare_to_do_exchange(fplog, re->destinations, replica_id, re->nrepl, &maxswap,
1314 re->order, re->cyclic, re->incycle, &bThisReplicaExchanged);
1316 /* Do intra-simulation broadcast so all processors belonging to
1317 * each simulation know whether they need to participate in
1318 * collecting the state. Otherwise, they might as well get on with
1319 * the next thing to do. */
1323 MPI_Bcast(&bThisReplicaExchanged, sizeof(gmx_bool), MPI_BYTE, MASTERRANK(cr),
1324 cr->mpi_comm_mygroup);
1328 if (bThisReplicaExchanged)
1330 /* Exchange the states */
1334 /* Collect the global state on the master node */
1335 if (DOMAINDECOMP(cr))
1337 dd_collect_state(cr->dd, state_local, state);
1341 pd_collect_state(cr, state);
1346 copy_state_nonatomdata(state_local, state);
1351 /* There will be only one swap cycle with standard replica
1352 * exchange, but there may be multiple swap cycles if we
1353 * allow multiple swaps. */
1355 for (j = 0; j < maxswap; j++)
1357 exchange_partner = re->order[replica_id][j];
1359 if (exchange_partner != replica_id)
1361 /* Exchange the global states between the master nodes */
1364 fprintf(debug, "Exchanging %d with %d\n", replica_id, exchange_partner);
1366 exchange_state(cr->ms, exchange_partner, state);
1369 /* For temperature-type replica exchange, we need to scale
1370 * the velocities. */
1371 if (re->type == ereTEMP || re->type == ereTL)
1373 scale_velocities(state, sqrt(re->q[ereTEMP][replica_id]/re->q[ereTEMP][re->destinations[replica_id]]));
1378 /* With domain decomposition the global state is distributed later */
1379 if (!DOMAINDECOMP(cr))
1381 /* Copy the global state to the local state data structure */
1382 copy_state_nonatomdata(state, state_local);
1386 bcast_state(cr, state, FALSE);
1391 return bThisReplicaExchanged;
1394 void print_replica_exchange_statistics(FILE *fplog, struct gmx_repl_ex *re)
1398 fprintf(fplog, "\nReplica exchange statistics\n");
1402 fprintf(fplog, "Repl %d attempts, %d odd, %d even\n",
1403 re->nattempt[0]+re->nattempt[1], re->nattempt[1], re->nattempt[0]);
1405 fprintf(fplog, "Repl average probabilities:\n");
1406 for (i = 1; i < re->nrepl; i++)
1408 if (re->nattempt[i%2] == 0)
1414 re->prob[i] = re->prob_sum[i]/re->nattempt[i%2];
1417 print_ind(fplog, "", re->nrepl, re->ind, NULL);
1418 print_prob(fplog, "", re->nrepl, re->prob);
1420 fprintf(fplog, "Repl number of exchanges:\n");
1421 print_ind(fplog, "", re->nrepl, re->ind, NULL);
1422 print_count(fplog, "", re->nrepl, re->nexchange);
1424 fprintf(fplog, "Repl average number of exchanges:\n");
1425 for (i = 1; i < re->nrepl; i++)
1427 if (re->nattempt[i%2] == 0)
1433 re->prob[i] = ((real)re->nexchange[i])/re->nattempt[i%2];
1436 print_ind(fplog, "", re->nrepl, re->ind, NULL);
1437 print_prob(fplog, "", re->nrepl, re->prob);
1439 fprintf(fplog, "\n");
1441 /* print the transition matrix */
1442 print_transition_matrix(fplog, re->nrepl, re->nmoves, re->nattempt);