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) 2013,2014,2015,2016,2017,2018,2019, 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.
39 * \brief This file defines integrators for energy minimization
41 * \author Berk Hess <hess@kth.se>
42 * \author Erik Lindahl <erik@kth.se>
43 * \ingroup module_mdrun
56 #include "gromacs/commandline/filenm.h"
57 #include "gromacs/domdec/collect.h"
58 #include "gromacs/domdec/dlbtiming.h"
59 #include "gromacs/domdec/domdec.h"
60 #include "gromacs/domdec/domdec_struct.h"
61 #include "gromacs/domdec/partition.h"
62 #include "gromacs/ewald/pme.h"
63 #include "gromacs/fileio/confio.h"
64 #include "gromacs/fileio/mtxio.h"
65 #include "gromacs/gmxlib/network.h"
66 #include "gromacs/gmxlib/nrnb.h"
67 #include "gromacs/imd/imd.h"
68 #include "gromacs/linearalgebra/sparsematrix.h"
69 #include "gromacs/listed_forces/manage_threading.h"
70 #include "gromacs/math/functions.h"
71 #include "gromacs/math/vec.h"
72 #include "gromacs/mdlib/constr.h"
73 #include "gromacs/mdlib/dispersioncorrection.h"
74 #include "gromacs/mdlib/ebin.h"
75 #include "gromacs/mdlib/energyoutput.h"
76 #include "gromacs/mdlib/force.h"
77 #include "gromacs/mdlib/forcerec.h"
78 #include "gromacs/mdlib/gmx_omp_nthreads.h"
79 #include "gromacs/mdlib/md_support.h"
80 #include "gromacs/mdlib/mdatoms.h"
81 #include "gromacs/mdlib/mdrun.h"
82 #include "gromacs/mdlib/mdsetup.h"
83 #include "gromacs/mdlib/ns.h"
84 #include "gromacs/mdlib/shellfc.h"
85 #include "gromacs/mdlib/sim_util.h"
86 #include "gromacs/mdlib/tgroup.h"
87 #include "gromacs/mdlib/trajectory_writing.h"
88 #include "gromacs/mdlib/update.h"
89 #include "gromacs/mdlib/vsite.h"
90 #include "gromacs/mdrunutility/printtime.h"
91 #include "gromacs/mdtypes/commrec.h"
92 #include "gromacs/mdtypes/inputrec.h"
93 #include "gromacs/mdtypes/md_enums.h"
94 #include "gromacs/mdtypes/state.h"
95 #include "gromacs/pbcutil/mshift.h"
96 #include "gromacs/pbcutil/pbc.h"
97 #include "gromacs/timing/wallcycle.h"
98 #include "gromacs/timing/walltime_accounting.h"
99 #include "gromacs/topology/mtop_util.h"
100 #include "gromacs/topology/topology.h"
101 #include "gromacs/utility/cstringutil.h"
102 #include "gromacs/utility/exceptions.h"
103 #include "gromacs/utility/fatalerror.h"
104 #include "gromacs/utility/logger.h"
105 #include "gromacs/utility/smalloc.h"
107 #include "integrator.h"
109 //! Utility structure for manipulating states during EM
111 //! Copy of the global state
114 PaddedVector<gmx::RVec> f;
117 //! Norm of the force
125 //! Print the EM starting conditions
126 static void print_em_start(FILE *fplog,
128 gmx_walltime_accounting_t walltime_accounting,
129 gmx_wallcycle_t wcycle,
132 walltime_accounting_start_time(walltime_accounting);
133 wallcycle_start(wcycle, ewcRUN);
134 print_start(fplog, cr, walltime_accounting, name);
137 //! Stop counting time for EM
138 static void em_time_end(gmx_walltime_accounting_t walltime_accounting,
139 gmx_wallcycle_t wcycle)
141 wallcycle_stop(wcycle, ewcRUN);
143 walltime_accounting_end_time(walltime_accounting);
146 //! Printing a log file and console header
147 static void sp_header(FILE *out, const char *minimizer, real ftol, int nsteps)
150 fprintf(out, "%s:\n", minimizer);
151 fprintf(out, " Tolerance (Fmax) = %12.5e\n", ftol);
152 fprintf(out, " Number of steps = %12d\n", nsteps);
155 //! Print warning message
156 static void warn_step(FILE *fp,
162 constexpr bool realIsDouble = GMX_DOUBLE;
165 if (!std::isfinite(fmax))
168 "\nEnergy minimization has stopped because the force "
169 "on at least one atom is not finite. This usually means "
170 "atoms are overlapping. Modify the input coordinates to "
171 "remove atom overlap or use soft-core potentials with "
172 "the free energy code to avoid infinite forces.\n%s",
174 "You could also be lucky that switching to double precision "
175 "is sufficient to obtain finite forces.\n" :
181 "\nEnergy minimization reached the maximum number "
182 "of steps before the forces reached the requested "
183 "precision Fmax < %g.\n", ftol);
188 "\nEnergy minimization has stopped, but the forces have "
189 "not converged to the requested precision Fmax < %g (which "
190 "may not be possible for your system). It stopped "
191 "because the algorithm tried to make a new step whose size "
192 "was too small, or there was no change in the energy since "
193 "last step. Either way, we regard the minimization as "
194 "converged to within the available machine precision, "
195 "given your starting configuration and EM parameters.\n%s%s",
198 "\nDouble precision normally gives you higher accuracy, but "
199 "this is often not needed for preparing to run molecular "
203 "You might need to increase your constraint accuracy, or turn\n"
204 "off constraints altogether (set constraints = none in mdp file)\n" :
208 fputs(wrap_lines(buffer, 78, 0, FALSE), stderr);
209 fputs(wrap_lines(buffer, 78, 0, FALSE), fp);
212 //! Print message about convergence of the EM
213 static void print_converged(FILE *fp, const char *alg, real ftol,
214 int64_t count, gmx_bool bDone, int64_t nsteps,
215 const em_state_t *ems, double sqrtNumAtoms)
217 char buf[STEPSTRSIZE];
221 fprintf(fp, "\n%s converged to Fmax < %g in %s steps\n",
222 alg, ftol, gmx_step_str(count, buf));
224 else if (count < nsteps)
226 fprintf(fp, "\n%s converged to machine precision in %s steps,\n"
227 "but did not reach the requested Fmax < %g.\n",
228 alg, gmx_step_str(count, buf), ftol);
232 fprintf(fp, "\n%s did not converge to Fmax < %g in %s steps.\n",
233 alg, ftol, gmx_step_str(count, buf));
237 fprintf(fp, "Potential Energy = %21.14e\n", ems->epot);
238 fprintf(fp, "Maximum force = %21.14e on atom %d\n", ems->fmax, ems->a_fmax + 1);
239 fprintf(fp, "Norm of force = %21.14e\n", ems->fnorm/sqrtNumAtoms);
241 fprintf(fp, "Potential Energy = %14.7e\n", ems->epot);
242 fprintf(fp, "Maximum force = %14.7e on atom %d\n", ems->fmax, ems->a_fmax + 1);
243 fprintf(fp, "Norm of force = %14.7e\n", ems->fnorm/sqrtNumAtoms);
247 //! Compute the norm and max of the force array in parallel
248 static void get_f_norm_max(const t_commrec *cr,
249 t_grpopts *opts, t_mdatoms *mdatoms, const rvec *f,
250 real *fnorm, real *fmax, int *a_fmax)
254 int la_max, a_max, start, end, i, m, gf;
256 /* This routine finds the largest force and returns it.
257 * On parallel machines the global max is taken.
263 end = mdatoms->homenr;
264 if (mdatoms->cFREEZE)
266 for (i = start; i < end; i++)
268 gf = mdatoms->cFREEZE[i];
270 for (m = 0; m < DIM; m++)
272 if (!opts->nFreeze[gf][m])
274 fam += gmx::square(f[i][m]);
287 for (i = start; i < end; i++)
299 if (la_max >= 0 && DOMAINDECOMP(cr))
301 a_max = cr->dd->globalAtomIndices[la_max];
309 snew(sum, 2*cr->nnodes+1);
310 sum[2*cr->nodeid] = fmax2;
311 sum[2*cr->nodeid+1] = a_max;
312 sum[2*cr->nnodes] = fnorm2;
313 gmx_sumd(2*cr->nnodes+1, sum, cr);
314 fnorm2 = sum[2*cr->nnodes];
315 /* Determine the global maximum */
316 for (i = 0; i < cr->nnodes; i++)
318 if (sum[2*i] > fmax2)
321 a_max = gmx::roundToInt(sum[2*i+1]);
329 *fnorm = sqrt(fnorm2);
341 //! Compute the norm of the force
342 static void get_state_f_norm_max(const t_commrec *cr,
343 t_grpopts *opts, t_mdatoms *mdatoms,
346 get_f_norm_max(cr, opts, mdatoms, ems->f.rvec_array(),
347 &ems->fnorm, &ems->fmax, &ems->a_fmax);
350 //! Initialize the energy minimization
351 static void init_em(FILE *fplog,
352 const gmx::MDLogger &mdlog,
355 const gmx_multisim_t *ms,
357 const MdrunOptions &mdrunOptions,
358 t_state *state_global, gmx_mtop_t *top_global,
359 em_state_t *ems, gmx_localtop_t *top,
362 t_graph **graph, gmx::MDAtoms *mdAtoms, gmx_global_stat_t *gstat,
363 gmx_vsite_t *vsite, gmx::Constraints *constr, gmx_shellfc_t **shellfc,
364 int nfile, const t_filenm fnm[])
370 fprintf(fplog, "Initiating %s\n", title);
375 state_global->ngtc = 0;
377 initialize_lambdas(fplog, *ir, MASTER(cr), &(state_global->fep_state), state_global->lambda, nullptr);
381 /* Interactive molecular dynamics */
382 init_IMD(ir, cr, ms, top_global, fplog, 1,
383 MASTER(cr) ? state_global->x.rvec_array() : nullptr,
384 nfile, fnm, nullptr, mdrunOptions);
388 GMX_ASSERT(shellfc != nullptr, "With NM we always support shells");
390 *shellfc = init_shell_flexcon(stdout,
392 constr ? constr->numFlexibleConstraints() : 0,
398 GMX_ASSERT(EI_ENERGY_MINIMIZATION(ir->eI), "This else currently only handles energy minimizers, consider if your algorithm needs shell/flexible-constraint support");
400 /* With energy minimization, shells and flexible constraints are
401 * automatically minimized when treated like normal DOFS.
403 if (shellfc != nullptr)
409 auto mdatoms = mdAtoms->mdatoms();
410 if (DOMAINDECOMP(cr))
412 top->useInDomainDecomp_ = true;
413 dd_init_local_top(*top_global, top);
415 dd_init_local_state(cr->dd, state_global, &ems->s);
417 /* Distribute the charge groups over the nodes from the master node */
418 dd_partition_system(fplog, mdlog, ir->init_step, cr, TRUE, 1,
419 state_global, *top_global, ir,
420 &ems->s, &ems->f, mdAtoms, top,
422 nrnb, nullptr, FALSE);
423 dd_store_state(cr->dd, &ems->s);
429 state_change_natoms(state_global, state_global->natoms);
430 /* Just copy the state */
431 ems->s = *state_global;
432 state_change_natoms(&ems->s, ems->s.natoms);
433 ems->f.resizeWithPadding(ems->s.natoms);
435 mdAlgorithmsSetupAtomData(cr, ir, *top_global, top, fr,
437 constr, vsite, shellfc ? *shellfc : nullptr);
441 set_vsite_top(vsite, top, mdatoms);
445 update_mdatoms(mdAtoms->mdatoms(), ems->s.lambda[efptMASS]);
449 // TODO how should this cross-module support dependency be managed?
450 if (ir->eConstrAlg == econtSHAKE &&
451 gmx_mtop_ftype_count(top_global, F_CONSTR) > 0)
453 gmx_fatal(FARGS, "Can not do energy minimization with %s, use %s\n",
454 econstr_names[econtSHAKE], econstr_names[econtLINCS]);
457 if (!ir->bContinuation)
459 /* Constrain the starting coordinates */
461 constr->apply(TRUE, TRUE,
463 ems->s.x.rvec_array(),
464 ems->s.x.rvec_array(),
467 ems->s.lambda[efptFEP], &dvdl_constr,
468 nullptr, nullptr, gmx::ConstraintVariable::Positions);
474 *gstat = global_stat_init(ir);
481 calc_shifts(ems->s.box, fr->shift_vec);
484 //! Finalize the minimization
485 static void finish_em(const t_commrec *cr, gmx_mdoutf_t outf,
486 gmx_walltime_accounting_t walltime_accounting,
487 gmx_wallcycle_t wcycle)
489 if (!thisRankHasDuty(cr, DUTY_PME))
491 /* Tell the PME only node to finish */
492 gmx_pme_send_finish(cr);
497 em_time_end(walltime_accounting, wcycle);
500 //! Swap two different EM states during minimization
501 static void swap_em_state(em_state_t **ems1, em_state_t **ems2)
510 //! Save the EM trajectory
511 static void write_em_traj(FILE *fplog, const t_commrec *cr,
513 gmx_bool bX, gmx_bool bF, const char *confout,
514 gmx_mtop_t *top_global,
515 t_inputrec *ir, int64_t step,
517 t_state *state_global,
518 ObservablesHistory *observablesHistory)
524 mdof_flags |= MDOF_X;
528 mdof_flags |= MDOF_F;
531 /* If we want IMD output, set appropriate MDOF flag */
534 mdof_flags |= MDOF_IMD;
537 mdoutf_write_to_trajectory_files(fplog, cr, outf, mdof_flags,
538 top_global, step, static_cast<double>(step),
539 &state->s, state_global, observablesHistory,
542 if (confout != nullptr)
544 if (DOMAINDECOMP(cr))
546 /* If bX=true, x was collected to state_global in the call above */
549 auto globalXRef = MASTER(cr) ? state_global->x : gmx::ArrayRef<gmx::RVec>();
550 dd_collect_vec(cr->dd, &state->s, state->s.x, globalXRef);
555 /* Copy the local state pointer */
556 state_global = &state->s;
561 if (ir->ePBC != epbcNONE && !ir->bPeriodicMols && DOMAINDECOMP(cr))
563 /* Make molecules whole only for confout writing */
564 do_pbc_mtop(ir->ePBC, state->s.box, top_global,
565 state_global->x.rvec_array());
568 write_sto_conf_mtop(confout,
569 *top_global->name, top_global,
570 state_global->x.rvec_array(), nullptr, ir->ePBC, state->s.box);
575 //! \brief Do one minimization step
577 // \returns true when the step succeeded, false when a constraint error occurred
578 static bool do_em_step(const t_commrec *cr,
579 t_inputrec *ir, t_mdatoms *md,
580 em_state_t *ems1, real a, const PaddedVector<gmx::RVec> *force,
582 gmx::Constraints *constr,
589 int nthreads gmx_unused;
591 bool validStep = true;
596 if (DOMAINDECOMP(cr) && s1->ddp_count != cr->dd->ddp_count)
598 gmx_incons("state mismatch in do_em_step");
601 s2->flags = s1->flags;
603 if (s2->natoms != s1->natoms)
605 state_change_natoms(s2, s1->natoms);
606 ems2->f.resizeWithPadding(s2->natoms);
608 if (DOMAINDECOMP(cr) && s2->cg_gl.size() != s1->cg_gl.size())
610 s2->cg_gl.resize(s1->cg_gl.size());
613 copy_mat(s1->box, s2->box);
614 /* Copy free energy state */
615 s2->lambda = s1->lambda;
616 copy_mat(s1->box, s2->box);
621 nthreads = gmx_omp_nthreads_get(emntUpdate);
622 #pragma omp parallel num_threads(nthreads)
624 const rvec *x1 = s1->x.rvec_array();
625 rvec *x2 = s2->x.rvec_array();
626 const rvec *f = force->rvec_array();
629 #pragma omp for schedule(static) nowait
630 for (int i = start; i < end; i++)
638 for (int m = 0; m < DIM; m++)
640 if (ir->opts.nFreeze[gf][m])
646 x2[i][m] = x1[i][m] + a*f[i][m];
650 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
653 if (s2->flags & (1<<estCGP))
655 /* Copy the CG p vector */
656 const rvec *p1 = s1->cg_p.rvec_array();
657 rvec *p2 = s2->cg_p.rvec_array();
658 #pragma omp for schedule(static) nowait
659 for (int i = start; i < end; i++)
661 // Trivial OpenMP block that does not throw
662 copy_rvec(p1[i], p2[i]);
666 if (DOMAINDECOMP(cr))
668 s2->ddp_count = s1->ddp_count;
670 /* OpenMP does not supported unsigned loop variables */
671 #pragma omp for schedule(static) nowait
672 for (int i = 0; i < gmx::ssize(s2->cg_gl); i++)
674 s2->cg_gl[i] = s1->cg_gl[i];
676 s2->ddp_count_cg_gl = s1->ddp_count_cg_gl;
684 constr->apply(TRUE, TRUE,
686 s1->x.rvec_array(), s2->x.rvec_array(),
688 s2->lambda[efptBONDED], &dvdl_constr,
689 nullptr, nullptr, gmx::ConstraintVariable::Positions);
693 /* This global reduction will affect performance at high
694 * parallelization, but we can not really avoid it.
695 * But usually EM is not run at high parallelization.
697 int reductionBuffer = static_cast<int>(!validStep);
698 gmx_sumi(1, &reductionBuffer, cr);
699 validStep = (reductionBuffer == 0);
702 // We should move this check to the different minimizers
703 if (!validStep && ir->eI != eiSteep)
705 gmx_fatal(FARGS, "The coordinates could not be constrained. Minimizer '%s' can not handle constraint failures, use minimizer '%s' before using '%s'.",
706 EI(ir->eI), EI(eiSteep), EI(ir->eI));
713 //! Prepare EM for using domain decomposition parallellization
714 static void em_dd_partition_system(FILE *fplog,
715 const gmx::MDLogger &mdlog,
716 int step, const t_commrec *cr,
717 gmx_mtop_t *top_global, t_inputrec *ir,
718 em_state_t *ems, gmx_localtop_t *top,
719 gmx::MDAtoms *mdAtoms, t_forcerec *fr,
720 gmx_vsite_t *vsite, gmx::Constraints *constr,
721 t_nrnb *nrnb, gmx_wallcycle_t wcycle)
723 /* Repartition the domain decomposition */
724 dd_partition_system(fplog, mdlog, step, cr, FALSE, 1,
725 nullptr, *top_global, ir,
727 mdAtoms, top, fr, vsite, constr,
728 nrnb, wcycle, FALSE);
729 dd_store_state(cr->dd, &ems->s);
735 /*! \brief Class to handle the work of setting and doing an energy evaluation.
737 * This class is a mere aggregate of parameters to pass to evaluate an
738 * energy, so that future changes to names and types of them consume
739 * less time when refactoring other code.
741 * Aggregate initialization is used, for which the chief risk is that
742 * if a member is added at the end and not all initializer lists are
743 * updated, then the member will be value initialized, which will
744 * typically mean initialization to zero.
746 * We only want to construct one of these with an initializer list, so
747 * we explicitly delete the default constructor. */
748 class EnergyEvaluator
751 //! We only intend to construct such objects with an initializer list.
752 EnergyEvaluator() = delete;
753 /*! \brief Evaluates an energy on the state in \c ems.
755 * \todo In practice, the same objects mu_tot, vir, and pres
756 * are always passed to this function, so we would rather have
757 * them as data members. However, their C-array types are
758 * unsuited for aggregate initialization. When the types
759 * improve, the call signature of this method can be reduced.
761 void run(em_state_t *ems, rvec mu_tot,
762 tensor vir, tensor pres,
763 int64_t count, gmx_bool bFirst);
764 //! Handles logging (deprecated).
767 const gmx::MDLogger &mdlog;
768 //! Handles communication.
770 //! Coordinates multi-simulations.
771 const gmx_multisim_t *ms;
772 //! Holds the simulation topology.
773 gmx_mtop_t *top_global;
774 //! Holds the domain topology.
776 //! User input options.
777 t_inputrec *inputrec;
778 //! Manages flop accounting.
780 //! Manages wall cycle accounting.
781 gmx_wallcycle_t wcycle;
782 //! Coordinates global reduction.
783 gmx_global_stat_t gstat;
784 //! Handles virtual sites.
786 //! Handles constraints.
787 gmx::Constraints *constr;
788 //! Handles strange things.
790 //! Molecular graph for SHAKE.
792 //! Per-atom data for this domain.
793 gmx::MDAtoms *mdAtoms;
794 //! Handles how to calculate the forces.
796 //! Schedule of force-calculation work each step for this task.
797 gmx::PpForceWorkload *ppForceWorkload;
798 //! Stores the computed energies.
799 gmx_enerdata_t *enerd;
803 EnergyEvaluator::run(em_state_t *ems, rvec mu_tot,
804 tensor vir, tensor pres,
805 int64_t count, gmx_bool bFirst)
809 tensor force_vir, shake_vir, ekin;
810 real dvdl_constr, prescorr, enercorr, dvdlcorr;
813 /* Set the time to the initial time, the time does not change during EM */
814 t = inputrec->init_t;
817 (DOMAINDECOMP(cr) && ems->s.ddp_count < cr->dd->ddp_count))
819 /* This is the first state or an old state used before the last ns */
825 if (inputrec->nstlist > 0)
833 construct_vsites(vsite, ems->s.x.rvec_array(), 1, nullptr,
834 top->idef.iparams, top->idef.il,
835 fr->ePBC, fr->bMolPBC, cr, ems->s.box);
838 if (DOMAINDECOMP(cr) && bNS)
840 /* Repartition the domain decomposition */
841 em_dd_partition_system(fplog, mdlog, count, cr, top_global, inputrec,
842 ems, top, mdAtoms, fr, vsite, constr,
846 /* Calc force & energy on new trial position */
847 /* do_force always puts the charge groups in the box and shifts again
848 * We do not unshift, so molecules are always whole in congrad.c
850 do_force(fplog, cr, ms, inputrec, nullptr, nullptr,
851 count, nrnb, wcycle, top, &top_global->groups,
852 ems->s.box, ems->s.x.arrayRefWithPadding(), &ems->s.hist,
853 ems->f.arrayRefWithPadding(), force_vir, mdAtoms->mdatoms(), enerd, fcd,
854 ems->s.lambda, graph, fr, ppForceWorkload, vsite, mu_tot, t, nullptr,
855 GMX_FORCE_STATECHANGED | GMX_FORCE_ALLFORCES |
856 GMX_FORCE_VIRIAL | GMX_FORCE_ENERGY |
857 (bNS ? GMX_FORCE_NS : 0),
858 DDBalanceRegionHandler(cr));
860 /* Clear the unused shake virial and pressure */
861 clear_mat(shake_vir);
864 /* Communicate stuff when parallel */
865 if (PAR(cr) && inputrec->eI != eiNM)
867 wallcycle_start(wcycle, ewcMoveE);
869 global_stat(gstat, cr, enerd, force_vir, shake_vir, mu_tot,
870 inputrec, nullptr, nullptr, nullptr, 1, &terminate,
876 wallcycle_stop(wcycle, ewcMoveE);
879 /* Calculate long range corrections to pressure and energy */
880 calc_dispcorr(inputrec, fr, ems->s.box, ems->s.lambda[efptVDW],
881 pres, force_vir, &prescorr, &enercorr, &dvdlcorr);
882 enerd->term[F_DISPCORR] = enercorr;
883 enerd->term[F_EPOT] += enercorr;
884 enerd->term[F_PRES] += prescorr;
885 enerd->term[F_DVDL] += dvdlcorr;
887 ems->epot = enerd->term[F_EPOT];
891 /* Project out the constraint components of the force */
893 rvec *f_rvec = ems->f.rvec_array();
894 constr->apply(FALSE, FALSE,
896 ems->s.x.rvec_array(), f_rvec, f_rvec,
898 ems->s.lambda[efptBONDED], &dvdl_constr,
899 nullptr, &shake_vir, gmx::ConstraintVariable::ForceDispl);
900 enerd->term[F_DVDL_CONSTR] += dvdl_constr;
901 m_add(force_vir, shake_vir, vir);
905 copy_mat(force_vir, vir);
909 enerd->term[F_PRES] =
910 calc_pres(fr->ePBC, inputrec->nwall, ems->s.box, ekin, vir, pres);
912 sum_dhdl(enerd, ems->s.lambda, inputrec->fepvals);
914 if (EI_ENERGY_MINIMIZATION(inputrec->eI))
916 get_state_f_norm_max(cr, &(inputrec->opts), mdAtoms->mdatoms(), ems);
922 //! Parallel utility summing energies and forces
923 static double reorder_partsum(const t_commrec *cr, t_grpopts *opts, t_mdatoms *mdatoms,
924 gmx_mtop_t *top_global,
925 em_state_t *s_min, em_state_t *s_b)
928 int ncg, *cg_gl, *index, c, cg, i, a0, a1, a, gf, m;
930 unsigned char *grpnrFREEZE;
934 fprintf(debug, "Doing reorder_partsum\n");
937 const rvec *fm = s_min->f.rvec_array();
938 const rvec *fb = s_b->f.rvec_array();
940 cgs_gl = dd_charge_groups_global(cr->dd);
941 index = cgs_gl->index;
943 /* Collect fm in a global vector fmg.
944 * This conflicts with the spirit of domain decomposition,
945 * but to fully optimize this a much more complicated algorithm is required.
948 snew(fmg, top_global->natoms);
950 ncg = s_min->s.cg_gl.size();
951 cg_gl = s_min->s.cg_gl.data();
953 for (c = 0; c < ncg; c++)
958 for (a = a0; a < a1; a++)
960 copy_rvec(fm[i], fmg[a]);
964 gmx_sum(top_global->natoms*3, fmg[0], cr);
966 /* Now we will determine the part of the sum for the cgs in state s_b */
967 ncg = s_b->s.cg_gl.size();
968 cg_gl = s_b->s.cg_gl.data();
972 grpnrFREEZE = top_global->groups.grpnr[egcFREEZE];
973 for (c = 0; c < ncg; c++)
978 for (a = a0; a < a1; a++)
980 if (mdatoms->cFREEZE && grpnrFREEZE)
984 for (m = 0; m < DIM; m++)
986 if (!opts->nFreeze[gf][m])
988 partsum += (fb[i][m] - fmg[a][m])*fb[i][m];
1000 //! Print some stuff, like beta, whatever that means.
1001 static real pr_beta(const t_commrec *cr, t_grpopts *opts, t_mdatoms *mdatoms,
1002 gmx_mtop_t *top_global,
1003 em_state_t *s_min, em_state_t *s_b)
1007 /* This is just the classical Polak-Ribiere calculation of beta;
1008 * it looks a bit complicated since we take freeze groups into account,
1009 * and might have to sum it in parallel runs.
1012 if (!DOMAINDECOMP(cr) ||
1013 (s_min->s.ddp_count == cr->dd->ddp_count &&
1014 s_b->s.ddp_count == cr->dd->ddp_count))
1016 const rvec *fm = s_min->f.rvec_array();
1017 const rvec *fb = s_b->f.rvec_array();
1020 /* This part of code can be incorrect with DD,
1021 * since the atom ordering in s_b and s_min might differ.
1023 for (int i = 0; i < mdatoms->homenr; i++)
1025 if (mdatoms->cFREEZE)
1027 gf = mdatoms->cFREEZE[i];
1029 for (int m = 0; m < DIM; m++)
1031 if (!opts->nFreeze[gf][m])
1033 sum += (fb[i][m] - fm[i][m])*fb[i][m];
1040 /* We need to reorder cgs while summing */
1041 sum = reorder_partsum(cr, opts, mdatoms, top_global, s_min, s_b);
1045 gmx_sumd(1, &sum, cr);
1048 return sum/gmx::square(s_min->fnorm);
1057 const char *CG = "Polak-Ribiere Conjugate Gradients";
1060 gmx_enerdata_t *enerd;
1061 gmx_global_stat_t gstat;
1063 double tmp, minstep;
1065 real a, b, c, beta = 0.0;
1068 gmx_bool converged, foundlower;
1070 gmx_bool do_log = FALSE, do_ene = FALSE, do_x, do_f;
1072 int number_steps, neval = 0, nstcg = inputrec->nstcgsteep;
1073 int m, step, nminstep;
1074 auto mdatoms = mdAtoms->mdatoms();
1076 GMX_LOG(mdlog.info).asParagraph().
1077 appendText("Note that activating conjugate gradient energy minimization via the "
1078 "integrator .mdp option and the command gmx mdrun may "
1079 "be available in a different form in a future version of GROMACS, "
1080 "e.g. gmx minimize and an .mdp option.");
1086 // In CG, the state is extended with a search direction
1087 state_global->flags |= (1<<estCGP);
1089 // Ensure the extra per-atom state array gets allocated
1090 state_change_natoms(state_global, state_global->natoms);
1092 // Initialize the search direction to zero
1093 for (RVec &cg_p : state_global->cg_p)
1099 /* Create 4 states on the stack and extract pointers that we will swap */
1100 em_state_t s0 {}, s1 {}, s2 {}, s3 {};
1101 em_state_t *s_min = &s0;
1102 em_state_t *s_a = &s1;
1103 em_state_t *s_b = &s2;
1104 em_state_t *s_c = &s3;
1106 /* Init em and store the local state in s_min */
1107 init_em(fplog, mdlog, CG, cr, ms, inputrec, mdrunOptions,
1108 state_global, top_global, s_min, &top,
1109 nrnb, fr, &graph, mdAtoms, &gstat,
1110 vsite, constr, nullptr,
1112 gmx_mdoutf *outf = init_mdoutf(fplog, nfile, fnm, mdrunOptions, cr, outputProvider, inputrec, top_global, nullptr, wcycle);
1114 init_enerdata(top_global->groups.grps[egcENER].nr, inputrec->fepvals->n_lambda, enerd);
1115 gmx::EnergyOutput energyOutput;
1116 energyOutput.prepare(mdoutf_get_fp_ene(outf), top_global, inputrec, nullptr);
1118 /* Print to log file */
1119 print_em_start(fplog, cr, walltime_accounting, wcycle, CG);
1121 /* Max number of steps */
1122 number_steps = inputrec->nsteps;
1126 sp_header(stderr, CG, inputrec->em_tol, number_steps);
1130 sp_header(fplog, CG, inputrec->em_tol, number_steps);
1133 EnergyEvaluator energyEvaluator {
1134 fplog, mdlog, cr, ms,
1136 inputrec, nrnb, wcycle, gstat,
1137 vsite, constr, fcd, graph,
1138 mdAtoms, fr, ppForceWorkload, enerd
1140 /* Call the force routine and some auxiliary (neighboursearching etc.) */
1141 /* do_force always puts the charge groups in the box and shifts again
1142 * We do not unshift, so molecules are always whole in congrad.c
1144 energyEvaluator.run(s_min, mu_tot, vir, pres, -1, TRUE);
1148 /* Copy stuff to the energy bin for easy printing etc. */
1149 matrix nullBox = {};
1150 energyOutput.addDataAtEnergyStep(false, false, static_cast<double>(step),
1151 mdatoms->tmass, enerd, nullptr, nullptr, nullptr, nullBox,
1152 nullptr, nullptr, vir, pres, nullptr, mu_tot, constr);
1154 print_ebin_header(fplog, step, step);
1155 energyOutput.printStepToEnergyFile(mdoutf_get_fp_ene(outf), TRUE, FALSE, FALSE,
1156 fplog, step, step, eprNORMAL,
1157 fcd, &(top_global->groups), &(inputrec->opts), nullptr);
1160 /* Estimate/guess the initial stepsize */
1161 stepsize = inputrec->em_stepsize/s_min->fnorm;
1165 double sqrtNumAtoms = sqrt(static_cast<double>(state_global->natoms));
1166 fprintf(stderr, " F-max = %12.5e on atom %d\n",
1167 s_min->fmax, s_min->a_fmax+1);
1168 fprintf(stderr, " F-Norm = %12.5e\n",
1169 s_min->fnorm/sqrtNumAtoms);
1170 fprintf(stderr, "\n");
1171 /* and copy to the log file too... */
1172 fprintf(fplog, " F-max = %12.5e on atom %d\n",
1173 s_min->fmax, s_min->a_fmax+1);
1174 fprintf(fplog, " F-Norm = %12.5e\n",
1175 s_min->fnorm/sqrtNumAtoms);
1176 fprintf(fplog, "\n");
1178 /* Start the loop over CG steps.
1179 * Each successful step is counted, and we continue until
1180 * we either converge or reach the max number of steps.
1183 for (step = 0; (number_steps < 0 || step <= number_steps) && !converged; step++)
1186 /* start taking steps in a new direction
1187 * First time we enter the routine, beta=0, and the direction is
1188 * simply the negative gradient.
1191 /* Calculate the new direction in p, and the gradient in this direction, gpa */
1192 rvec *pm = s_min->s.cg_p.rvec_array();
1193 const rvec *sfm = s_min->f.rvec_array();
1196 for (int i = 0; i < mdatoms->homenr; i++)
1198 if (mdatoms->cFREEZE)
1200 gf = mdatoms->cFREEZE[i];
1202 for (m = 0; m < DIM; m++)
1204 if (!inputrec->opts.nFreeze[gf][m])
1206 pm[i][m] = sfm[i][m] + beta*pm[i][m];
1207 gpa -= pm[i][m]*sfm[i][m];
1208 /* f is negative gradient, thus the sign */
1217 /* Sum the gradient along the line across CPUs */
1220 gmx_sumd(1, &gpa, cr);
1223 /* Calculate the norm of the search vector */
1224 get_f_norm_max(cr, &(inputrec->opts), mdatoms, pm, &pnorm, nullptr, nullptr);
1226 /* Just in case stepsize reaches zero due to numerical precision... */
1229 stepsize = inputrec->em_stepsize/pnorm;
1233 * Double check the value of the derivative in the search direction.
1234 * If it is positive it must be due to the old information in the
1235 * CG formula, so just remove that and start over with beta=0.
1236 * This corresponds to a steepest descent step.
1241 step--; /* Don't count this step since we are restarting */
1242 continue; /* Go back to the beginning of the big for-loop */
1245 /* Calculate minimum allowed stepsize, before the average (norm)
1246 * relative change in coordinate is smaller than precision
1249 auto s_min_x = makeArrayRef(s_min->s.x);
1250 for (int i = 0; i < mdatoms->homenr; i++)
1252 for (m = 0; m < DIM; m++)
1254 tmp = fabs(s_min_x[i][m]);
1263 /* Add up from all CPUs */
1266 gmx_sumd(1, &minstep, cr);
1269 minstep = GMX_REAL_EPS/sqrt(minstep/(3*top_global->natoms));
1271 if (stepsize < minstep)
1277 /* Write coordinates if necessary */
1278 do_x = do_per_step(step, inputrec->nstxout);
1279 do_f = do_per_step(step, inputrec->nstfout);
1281 write_em_traj(fplog, cr, outf, do_x, do_f, nullptr,
1282 top_global, inputrec, step,
1283 s_min, state_global, observablesHistory);
1285 /* Take a step downhill.
1286 * In theory, we should minimize the function along this direction.
1287 * That is quite possible, but it turns out to take 5-10 function evaluations
1288 * for each line. However, we dont really need to find the exact minimum -
1289 * it is much better to start a new CG step in a modified direction as soon
1290 * as we are close to it. This will save a lot of energy evaluations.
1292 * In practice, we just try to take a single step.
1293 * If it worked (i.e. lowered the energy), we increase the stepsize but
1294 * the continue straight to the next CG step without trying to find any minimum.
1295 * If it didn't work (higher energy), there must be a minimum somewhere between
1296 * the old position and the new one.
1298 * Due to the finite numerical accuracy, it turns out that it is a good idea
1299 * to even accept a SMALL increase in energy, if the derivative is still downhill.
1300 * This leads to lower final energies in the tests I've done. / Erik
1302 s_a->epot = s_min->epot;
1304 c = a + stepsize; /* reference position along line is zero */
1306 if (DOMAINDECOMP(cr) && s_min->s.ddp_count < cr->dd->ddp_count)
1308 em_dd_partition_system(fplog, mdlog, step, cr, top_global, inputrec,
1309 s_min, &top, mdAtoms, fr, vsite, constr,
1313 /* Take a trial step (new coords in s_c) */
1314 do_em_step(cr, inputrec, mdatoms, s_min, c, &s_min->s.cg_p, s_c,
1318 /* Calculate energy for the trial step */
1319 energyEvaluator.run(s_c, mu_tot, vir, pres, -1, FALSE);
1321 /* Calc derivative along line */
1322 const rvec *pc = s_c->s.cg_p.rvec_array();
1323 const rvec *sfc = s_c->f.rvec_array();
1325 for (int i = 0; i < mdatoms->homenr; i++)
1327 for (m = 0; m < DIM; m++)
1329 gpc -= pc[i][m]*sfc[i][m]; /* f is negative gradient, thus the sign */
1332 /* Sum the gradient along the line across CPUs */
1335 gmx_sumd(1, &gpc, cr);
1338 /* This is the max amount of increase in energy we tolerate */
1339 tmp = std::sqrt(GMX_REAL_EPS)*fabs(s_a->epot);
1341 /* Accept the step if the energy is lower, or if it is not significantly higher
1342 * and the line derivative is still negative.
1344 if (s_c->epot < s_a->epot || (gpc < 0 && s_c->epot < (s_a->epot + tmp)))
1347 /* Great, we found a better energy. Increase step for next iteration
1348 * if we are still going down, decrease it otherwise
1352 stepsize *= 1.618034; /* The golden section */
1356 stepsize *= 0.618034; /* 1/golden section */
1361 /* New energy is the same or higher. We will have to do some work
1362 * to find a smaller value in the interval. Take smaller step next time!
1365 stepsize *= 0.618034;
1371 /* OK, if we didn't find a lower value we will have to locate one now - there must
1372 * be one in the interval [a=0,c].
1373 * The same thing is valid here, though: Don't spend dozens of iterations to find
1374 * the line minimum. We try to interpolate based on the derivative at the endpoints,
1375 * and only continue until we find a lower value. In most cases this means 1-2 iterations.
1377 * I also have a safeguard for potentially really pathological functions so we never
1378 * take more than 20 steps before we give up ...
1380 * If we already found a lower value we just skip this step and continue to the update.
1389 /* Select a new trial point.
1390 * If the derivatives at points a & c have different sign we interpolate to zero,
1391 * otherwise just do a bisection.
1393 if (gpa < 0 && gpc > 0)
1395 b = a + gpa*(a-c)/(gpc-gpa);
1402 /* safeguard if interpolation close to machine accuracy causes errors:
1403 * never go outside the interval
1405 if (b <= a || b >= c)
1410 if (DOMAINDECOMP(cr) && s_min->s.ddp_count != cr->dd->ddp_count)
1412 /* Reload the old state */
1413 em_dd_partition_system(fplog, mdlog, -1, cr, top_global, inputrec,
1414 s_min, &top, mdAtoms, fr, vsite, constr,
1418 /* Take a trial step to this new point - new coords in s_b */
1419 do_em_step(cr, inputrec, mdatoms, s_min, b, &s_min->s.cg_p, s_b,
1423 /* Calculate energy for the trial step */
1424 energyEvaluator.run(s_b, mu_tot, vir, pres, -1, FALSE);
1426 /* p does not change within a step, but since the domain decomposition
1427 * might change, we have to use cg_p of s_b here.
1429 const rvec *pb = s_b->s.cg_p.rvec_array();
1430 const rvec *sfb = s_b->f.rvec_array();
1432 for (int i = 0; i < mdatoms->homenr; i++)
1434 for (m = 0; m < DIM; m++)
1436 gpb -= pb[i][m]*sfb[i][m]; /* f is negative gradient, thus the sign */
1439 /* Sum the gradient along the line across CPUs */
1442 gmx_sumd(1, &gpb, cr);
1447 fprintf(debug, "CGE: EpotA %f EpotB %f EpotC %f gpb %f\n",
1448 s_a->epot, s_b->epot, s_c->epot, gpb);
1451 epot_repl = s_b->epot;
1453 /* Keep one of the intervals based on the value of the derivative at the new point */
1456 /* Replace c endpoint with b */
1457 swap_em_state(&s_b, &s_c);
1463 /* Replace a endpoint with b */
1464 swap_em_state(&s_b, &s_a);
1470 * Stop search as soon as we find a value smaller than the endpoints.
1471 * Never run more than 20 steps, no matter what.
1475 while ((epot_repl > s_a->epot || epot_repl > s_c->epot) &&
1478 if (std::fabs(epot_repl - s_min->epot) < fabs(s_min->epot)*GMX_REAL_EPS ||
1481 /* OK. We couldn't find a significantly lower energy.
1482 * If beta==0 this was steepest descent, and then we give up.
1483 * If not, set beta=0 and restart with steepest descent before quitting.
1493 /* Reset memory before giving up */
1499 /* Select min energy state of A & C, put the best in B.
1501 if (s_c->epot < s_a->epot)
1505 fprintf(debug, "CGE: C (%f) is lower than A (%f), moving C to B\n",
1506 s_c->epot, s_a->epot);
1508 swap_em_state(&s_b, &s_c);
1515 fprintf(debug, "CGE: A (%f) is lower than C (%f), moving A to B\n",
1516 s_a->epot, s_c->epot);
1518 swap_em_state(&s_b, &s_a);
1527 fprintf(debug, "CGE: Found a lower energy %f, moving C to B\n",
1530 swap_em_state(&s_b, &s_c);
1534 /* new search direction */
1535 /* beta = 0 means forget all memory and restart with steepest descents. */
1536 if (nstcg && ((step % nstcg) == 0))
1542 /* s_min->fnorm cannot be zero, because then we would have converged
1546 /* Polak-Ribiere update.
1547 * Change to fnorm2/fnorm2_old for Fletcher-Reeves
1549 beta = pr_beta(cr, &inputrec->opts, mdatoms, top_global, s_min, s_b);
1551 /* Limit beta to prevent oscillations */
1552 if (fabs(beta) > 5.0)
1558 /* update positions */
1559 swap_em_state(&s_min, &s_b);
1562 /* Print it if necessary */
1565 if (mdrunOptions.verbose)
1567 double sqrtNumAtoms = sqrt(static_cast<double>(state_global->natoms));
1568 fprintf(stderr, "\rStep %d, Epot=%12.6e, Fnorm=%9.3e, Fmax=%9.3e (atom %d)\n",
1569 step, s_min->epot, s_min->fnorm/sqrtNumAtoms,
1570 s_min->fmax, s_min->a_fmax+1);
1573 /* Store the new (lower) energies */
1574 matrix nullBox = {};
1575 energyOutput.addDataAtEnergyStep(false, false, static_cast<double>(step),
1576 mdatoms->tmass, enerd, nullptr, nullptr, nullptr, nullBox,
1577 nullptr, nullptr, vir, pres, nullptr, mu_tot, constr);
1579 do_log = do_per_step(step, inputrec->nstlog);
1580 do_ene = do_per_step(step, inputrec->nstenergy);
1582 /* Prepare IMD energy record, if bIMD is TRUE. */
1583 IMD_fill_energy_record(inputrec->bIMD, inputrec->imd, enerd, step, TRUE);
1587 print_ebin_header(fplog, step, step);
1589 energyOutput.printStepToEnergyFile(mdoutf_get_fp_ene(outf), do_ene, FALSE, FALSE,
1590 do_log ? fplog : nullptr, step, step, eprNORMAL,
1591 fcd, &(top_global->groups), &(inputrec->opts), nullptr);
1594 /* Send energies and positions to the IMD client if bIMD is TRUE. */
1595 if (MASTER(cr) && do_IMD(inputrec->bIMD, step, cr, TRUE, state_global->box, state_global->x.rvec_array(), inputrec, 0, wcycle))
1597 IMD_send_positions(inputrec->imd);
1600 /* Stop when the maximum force lies below tolerance.
1601 * If we have reached machine precision, converged is already set to true.
1603 converged = converged || (s_min->fmax < inputrec->em_tol);
1605 } /* End of the loop */
1607 /* IMD cleanup, if bIMD is TRUE. */
1608 IMD_finalize(inputrec->bIMD, inputrec->imd);
1612 step--; /* we never took that last step in this case */
1615 if (s_min->fmax > inputrec->em_tol)
1619 warn_step(fplog, inputrec->em_tol, s_min->fmax,
1620 step-1 == number_steps, FALSE);
1627 /* If we printed energy and/or logfile last step (which was the last step)
1628 * we don't have to do it again, but otherwise print the final values.
1632 /* Write final value to log since we didn't do anything the last step */
1633 print_ebin_header(fplog, step, step);
1635 if (!do_ene || !do_log)
1637 /* Write final energy file entries */
1638 energyOutput.printStepToEnergyFile(mdoutf_get_fp_ene(outf), !do_ene, FALSE, FALSE,
1639 !do_log ? fplog : nullptr, step, step, eprNORMAL,
1640 fcd, &(top_global->groups), &(inputrec->opts), nullptr);
1644 /* Print some stuff... */
1647 fprintf(stderr, "\nwriting lowest energy coordinates.\n");
1651 * For accurate normal mode calculation it is imperative that we
1652 * store the last conformation into the full precision binary trajectory.
1654 * However, we should only do it if we did NOT already write this step
1655 * above (which we did if do_x or do_f was true).
1657 /* Note that with 0 < nstfout != nstxout we can end up with two frames
1658 * in the trajectory with the same step number.
1660 do_x = !do_per_step(step, inputrec->nstxout);
1661 do_f = (inputrec->nstfout > 0 && !do_per_step(step, inputrec->nstfout));
1663 write_em_traj(fplog, cr, outf, do_x, do_f, ftp2fn(efSTO, nfile, fnm),
1664 top_global, inputrec, step,
1665 s_min, state_global, observablesHistory);
1670 double sqrtNumAtoms = sqrt(static_cast<double>(state_global->natoms));
1671 print_converged(stderr, CG, inputrec->em_tol, step, converged, number_steps,
1672 s_min, sqrtNumAtoms);
1673 print_converged(fplog, CG, inputrec->em_tol, step, converged, number_steps,
1674 s_min, sqrtNumAtoms);
1676 fprintf(fplog, "\nPerformed %d energy evaluations in total.\n", neval);
1679 finish_em(cr, outf, walltime_accounting, wcycle);
1681 /* To print the actual number of steps we needed somewhere */
1682 walltime_accounting_set_nsteps_done(walltime_accounting, step);
1687 Integrator::do_lbfgs()
1689 static const char *LBFGS = "Low-Memory BFGS Minimizer";
1692 gmx_enerdata_t *enerd;
1693 gmx_global_stat_t gstat;
1695 int ncorr, nmaxcorr, point, cp, neval, nminstep;
1696 double stepsize, step_taken, gpa, gpb, gpc, tmp, minstep;
1697 real *rho, *alpha, *p, *s, **dx, **dg;
1698 real a, b, c, maxdelta, delta;
1700 real dgdx, dgdg, sq, yr, beta;
1703 gmx_bool do_log, do_ene, do_x, do_f, foundlower, *frozen;
1705 int start, end, number_steps;
1706 int i, k, m, n, gf, step;
1708 auto mdatoms = mdAtoms->mdatoms();
1710 GMX_LOG(mdlog.info).asParagraph().
1711 appendText("Note that activating L-BFGS energy minimization via the "
1712 "integrator .mdp option and the command gmx mdrun may "
1713 "be available in a different form in a future version of GROMACS, "
1714 "e.g. gmx minimize and an .mdp option.");
1718 gmx_fatal(FARGS, "L-BFGS minimization only supports a single rank");
1721 if (nullptr != constr)
1723 gmx_fatal(FARGS, "The combination of constraints and L-BFGS minimization is not implemented. Either do not use constraints, or use another minimizer (e.g. steepest descent).");
1726 n = 3*state_global->natoms;
1727 nmaxcorr = inputrec->nbfgscorr;
1732 snew(rho, nmaxcorr);
1733 snew(alpha, nmaxcorr);
1736 for (i = 0; i < nmaxcorr; i++)
1742 for (i = 0; i < nmaxcorr; i++)
1751 init_em(fplog, mdlog, LBFGS, cr, ms, inputrec, mdrunOptions,
1752 state_global, top_global, &ems, &top,
1753 nrnb, fr, &graph, mdAtoms, &gstat,
1754 vsite, constr, nullptr,
1756 gmx_mdoutf *outf = init_mdoutf(fplog, nfile, fnm, mdrunOptions, cr, outputProvider, inputrec, top_global, nullptr, wcycle);
1758 init_enerdata(top_global->groups.grps[egcENER].nr, inputrec->fepvals->n_lambda, enerd);
1759 gmx::EnergyOutput energyOutput;
1760 energyOutput.prepare(mdoutf_get_fp_ene(outf), top_global, inputrec, nullptr);
1763 end = mdatoms->homenr;
1765 /* We need 4 working states */
1766 em_state_t s0 {}, s1 {}, s2 {}, s3 {};
1767 em_state_t *sa = &s0;
1768 em_state_t *sb = &s1;
1769 em_state_t *sc = &s2;
1770 em_state_t *last = &s3;
1771 /* Initialize by copying the state from ems (we could skip x and f here) */
1776 /* Print to log file */
1777 print_em_start(fplog, cr, walltime_accounting, wcycle, LBFGS);
1779 do_log = do_ene = do_x = do_f = TRUE;
1781 /* Max number of steps */
1782 number_steps = inputrec->nsteps;
1784 /* Create a 3*natoms index to tell whether each degree of freedom is frozen */
1786 for (i = start; i < end; i++)
1788 if (mdatoms->cFREEZE)
1790 gf = mdatoms->cFREEZE[i];
1792 for (m = 0; m < DIM; m++)
1794 frozen[3*i+m] = (inputrec->opts.nFreeze[gf][m] != 0);
1799 sp_header(stderr, LBFGS, inputrec->em_tol, number_steps);
1803 sp_header(fplog, LBFGS, inputrec->em_tol, number_steps);
1808 construct_vsites(vsite, state_global->x.rvec_array(), 1, nullptr,
1809 top.idef.iparams, top.idef.il,
1810 fr->ePBC, fr->bMolPBC, cr, state_global->box);
1813 /* Call the force routine and some auxiliary (neighboursearching etc.) */
1814 /* do_force always puts the charge groups in the box and shifts again
1815 * We do not unshift, so molecules are always whole
1818 EnergyEvaluator energyEvaluator {
1819 fplog, mdlog, cr, ms,
1821 inputrec, nrnb, wcycle, gstat,
1822 vsite, constr, fcd, graph,
1823 mdAtoms, fr, ppForceWorkload, enerd
1825 energyEvaluator.run(&ems, mu_tot, vir, pres, -1, TRUE);
1829 /* Copy stuff to the energy bin for easy printing etc. */
1830 matrix nullBox = {};
1831 energyOutput.addDataAtEnergyStep(false, false, static_cast<double>(step),
1832 mdatoms->tmass, enerd, nullptr, nullptr, nullptr, nullBox,
1833 nullptr, nullptr, vir, pres, nullptr, mu_tot, constr);
1835 print_ebin_header(fplog, step, step);
1836 energyOutput.printStepToEnergyFile(mdoutf_get_fp_ene(outf), TRUE, FALSE, FALSE,
1837 fplog, step, step, eprNORMAL,
1838 fcd, &(top_global->groups), &(inputrec->opts), nullptr);
1841 /* Set the initial step.
1842 * since it will be multiplied by the non-normalized search direction
1843 * vector (force vector the first time), we scale it by the
1844 * norm of the force.
1849 double sqrtNumAtoms = sqrt(static_cast<double>(state_global->natoms));
1850 fprintf(stderr, "Using %d BFGS correction steps.\n\n", nmaxcorr);
1851 fprintf(stderr, " F-max = %12.5e on atom %d\n", ems.fmax, ems.a_fmax + 1);
1852 fprintf(stderr, " F-Norm = %12.5e\n", ems.fnorm/sqrtNumAtoms);
1853 fprintf(stderr, "\n");
1854 /* and copy to the log file too... */
1855 fprintf(fplog, "Using %d BFGS correction steps.\n\n", nmaxcorr);
1856 fprintf(fplog, " F-max = %12.5e on atom %d\n", ems.fmax, ems.a_fmax + 1);
1857 fprintf(fplog, " F-Norm = %12.5e\n", ems.fnorm/sqrtNumAtoms);
1858 fprintf(fplog, "\n");
1861 // Point is an index to the memory of search directions, where 0 is the first one.
1864 // Set initial search direction to the force (-gradient), or 0 for frozen particles.
1865 real *fInit = static_cast<real *>(ems.f.rvec_array()[0]);
1866 for (i = 0; i < n; i++)
1870 dx[point][i] = fInit[i]; /* Initial search direction */
1878 // Stepsize will be modified during the search, and actually it is not critical
1879 // (the main efficiency in the algorithm comes from changing directions), but
1880 // we still need an initial value, so estimate it as the inverse of the norm
1881 // so we take small steps where the potential fluctuates a lot.
1882 stepsize = 1.0/ems.fnorm;
1884 /* Start the loop over BFGS steps.
1885 * Each successful step is counted, and we continue until
1886 * we either converge or reach the max number of steps.
1891 /* Set the gradient from the force */
1893 for (step = 0; (number_steps < 0 || step <= number_steps) && !converged; step++)
1896 /* Write coordinates if necessary */
1897 do_x = do_per_step(step, inputrec->nstxout);
1898 do_f = do_per_step(step, inputrec->nstfout);
1903 mdof_flags |= MDOF_X;
1908 mdof_flags |= MDOF_F;
1913 mdof_flags |= MDOF_IMD;
1916 mdoutf_write_to_trajectory_files(fplog, cr, outf, mdof_flags,
1917 top_global, step, static_cast<real>(step), &ems.s, state_global, observablesHistory, ems.f);
1919 /* Do the linesearching in the direction dx[point][0..(n-1)] */
1921 /* make s a pointer to current search direction - point=0 first time we get here */
1924 real *xx = static_cast<real *>(ems.s.x.rvec_array()[0]);
1925 real *ff = static_cast<real *>(ems.f.rvec_array()[0]);
1927 // calculate line gradient in position A
1928 for (gpa = 0, i = 0; i < n; i++)
1933 /* Calculate minimum allowed stepsize along the line, before the average (norm)
1934 * relative change in coordinate is smaller than precision
1936 for (minstep = 0, i = 0; i < n; i++)
1946 minstep = GMX_REAL_EPS/sqrt(minstep/n);
1948 if (stepsize < minstep)
1954 // Before taking any steps along the line, store the old position
1956 real *lastx = static_cast<real *>(last->s.x.data()[0]);
1957 real *lastf = static_cast<real *>(last->f.data()[0]);
1962 /* Take a step downhill.
1963 * In theory, we should find the actual minimum of the function in this
1964 * direction, somewhere along the line.
1965 * That is quite possible, but it turns out to take 5-10 function evaluations
1966 * for each line. However, we dont really need to find the exact minimum -
1967 * it is much better to start a new BFGS step in a modified direction as soon
1968 * as we are close to it. This will save a lot of energy evaluations.
1970 * In practice, we just try to take a single step.
1971 * If it worked (i.e. lowered the energy), we increase the stepsize but
1972 * continue straight to the next BFGS step without trying to find any minimum,
1973 * i.e. we change the search direction too. If the line was smooth, it is
1974 * likely we are in a smooth region, and then it makes sense to take longer
1975 * steps in the modified search direction too.
1977 * If it didn't work (higher energy), there must be a minimum somewhere between
1978 * the old position and the new one. Then we need to start by finding a lower
1979 * value before we change search direction. Since the energy was apparently
1980 * quite rough, we need to decrease the step size.
1982 * Due to the finite numerical accuracy, it turns out that it is a good idea
1983 * to accept a SMALL increase in energy, if the derivative is still downhill.
1984 * This leads to lower final energies in the tests I've done. / Erik
1987 // State "A" is the first position along the line.
1988 // reference position along line is initially zero
1991 // Check stepsize first. We do not allow displacements
1992 // larger than emstep.
1996 // Pick a new position C by adding stepsize to A.
1999 // Calculate what the largest change in any individual coordinate
2000 // would be (translation along line * gradient along line)
2002 for (i = 0; i < n; i++)
2005 if (delta > maxdelta)
2010 // If any displacement is larger than the stepsize limit, reduce the step
2011 if (maxdelta > inputrec->em_stepsize)
2016 while (maxdelta > inputrec->em_stepsize);
2018 // Take a trial step and move the coordinate array xc[] to position C
2019 real *xc = static_cast<real *>(sc->s.x.rvec_array()[0]);
2020 for (i = 0; i < n; i++)
2022 xc[i] = lastx[i] + c*s[i];
2026 // Calculate energy for the trial step in position C
2027 energyEvaluator.run(sc, mu_tot, vir, pres, step, FALSE);
2029 // Calc line gradient in position C
2030 real *fc = static_cast<real *>(sc->f.rvec_array()[0]);
2031 for (gpc = 0, i = 0; i < n; i++)
2033 gpc -= s[i]*fc[i]; /* f is negative gradient, thus the sign */
2035 /* Sum the gradient along the line across CPUs */
2038 gmx_sumd(1, &gpc, cr);
2041 // This is the max amount of increase in energy we tolerate.
2042 // By allowing VERY small changes (close to numerical precision) we
2043 // frequently find even better (lower) final energies.
2044 tmp = std::sqrt(GMX_REAL_EPS)*fabs(sa->epot);
2046 // Accept the step if the energy is lower in the new position C (compared to A),
2047 // or if it is not significantly higher and the line derivative is still negative.
2048 foundlower = sc->epot < sa->epot || (gpc < 0 && sc->epot < (sa->epot + tmp));
2049 // If true, great, we found a better energy. We no longer try to alter the
2050 // stepsize, but simply accept this new better position. The we select a new
2051 // search direction instead, which will be much more efficient than continuing
2052 // to take smaller steps along a line. Set fnorm based on the new C position,
2053 // which will be used to update the stepsize to 1/fnorm further down.
2055 // If false, the energy is NOT lower in point C, i.e. it will be the same
2056 // or higher than in point A. In this case it is pointless to move to point C,
2057 // so we will have to do more iterations along the same line to find a smaller
2058 // value in the interval [A=0.0,C].
2059 // Here, A is still 0.0, but that will change when we do a search in the interval
2060 // [0.0,C] below. That search we will do by interpolation or bisection rather
2061 // than with the stepsize, so no need to modify it. For the next search direction
2062 // it will be reset to 1/fnorm anyway.
2066 // OK, if we didn't find a lower value we will have to locate one now - there must
2067 // be one in the interval [a,c].
2068 // The same thing is valid here, though: Don't spend dozens of iterations to find
2069 // the line minimum. We try to interpolate based on the derivative at the endpoints,
2070 // and only continue until we find a lower value. In most cases this means 1-2 iterations.
2071 // I also have a safeguard for potentially really pathological functions so we never
2072 // take more than 20 steps before we give up.
2073 // If we already found a lower value we just skip this step and continue to the update.
2078 // Select a new trial point B in the interval [A,C].
2079 // If the derivatives at points a & c have different sign we interpolate to zero,
2080 // otherwise just do a bisection since there might be multiple minima/maxima
2081 // inside the interval.
2082 if (gpa < 0 && gpc > 0)
2084 b = a + gpa*(a-c)/(gpc-gpa);
2091 /* safeguard if interpolation close to machine accuracy causes errors:
2092 * never go outside the interval
2094 if (b <= a || b >= c)
2099 // Take a trial step to point B
2100 real *xb = static_cast<real *>(sb->s.x.rvec_array()[0]);
2101 for (i = 0; i < n; i++)
2103 xb[i] = lastx[i] + b*s[i];
2107 // Calculate energy for the trial step in point B
2108 energyEvaluator.run(sb, mu_tot, vir, pres, step, FALSE);
2111 // Calculate gradient in point B
2112 real *fb = static_cast<real *>(sb->f.rvec_array()[0]);
2113 for (gpb = 0, i = 0; i < n; i++)
2115 gpb -= s[i]*fb[i]; /* f is negative gradient, thus the sign */
2118 /* Sum the gradient along the line across CPUs */
2121 gmx_sumd(1, &gpb, cr);
2124 // Keep one of the intervals [A,B] or [B,C] based on the value of the derivative
2125 // at the new point B, and rename the endpoints of this new interval A and C.
2128 /* Replace c endpoint with b */
2130 /* swap states b and c */
2131 swap_em_state(&sb, &sc);
2135 /* Replace a endpoint with b */
2137 /* swap states a and b */
2138 swap_em_state(&sa, &sb);
2142 * Stop search as soon as we find a value smaller than the endpoints,
2143 * or if the tolerance is below machine precision.
2144 * Never run more than 20 steps, no matter what.
2148 while ((sb->epot > sa->epot || sb->epot > sc->epot) && (nminstep < 20));
2150 if (std::fabs(sb->epot - Epot0) < GMX_REAL_EPS || nminstep >= 20)
2152 /* OK. We couldn't find a significantly lower energy.
2153 * If ncorr==0 this was steepest descent, and then we give up.
2154 * If not, reset memory to restart as steepest descent before quitting.
2166 /* Search in gradient direction */
2167 for (i = 0; i < n; i++)
2169 dx[point][i] = ff[i];
2171 /* Reset stepsize */
2172 stepsize = 1.0/fnorm;
2177 /* Select min energy state of A & C, put the best in xx/ff/Epot
2179 if (sc->epot < sa->epot)
2201 /* Update the memory information, and calculate a new
2202 * approximation of the inverse hessian
2205 /* Have new data in Epot, xx, ff */
2206 if (ncorr < nmaxcorr)
2211 for (i = 0; i < n; i++)
2213 dg[point][i] = lastf[i]-ff[i];
2214 dx[point][i] *= step_taken;
2219 for (i = 0; i < n; i++)
2221 dgdg += dg[point][i]*dg[point][i];
2222 dgdx += dg[point][i]*dx[point][i];
2227 rho[point] = 1.0/dgdx;
2230 if (point >= nmaxcorr)
2236 for (i = 0; i < n; i++)
2243 /* Recursive update. First go back over the memory points */
2244 for (k = 0; k < ncorr; k++)
2253 for (i = 0; i < n; i++)
2255 sq += dx[cp][i]*p[i];
2258 alpha[cp] = rho[cp]*sq;
2260 for (i = 0; i < n; i++)
2262 p[i] -= alpha[cp]*dg[cp][i];
2266 for (i = 0; i < n; i++)
2271 /* And then go forward again */
2272 for (k = 0; k < ncorr; k++)
2275 for (i = 0; i < n; i++)
2277 yr += p[i]*dg[cp][i];
2281 beta = alpha[cp]-beta;
2283 for (i = 0; i < n; i++)
2285 p[i] += beta*dx[cp][i];
2295 for (i = 0; i < n; i++)
2299 dx[point][i] = p[i];
2307 /* Print it if necessary */
2310 if (mdrunOptions.verbose)
2312 double sqrtNumAtoms = sqrt(static_cast<double>(state_global->natoms));
2313 fprintf(stderr, "\rStep %d, Epot=%12.6e, Fnorm=%9.3e, Fmax=%9.3e (atom %d)\n",
2314 step, ems.epot, ems.fnorm/sqrtNumAtoms, ems.fmax, ems.a_fmax + 1);
2317 /* Store the new (lower) energies */
2318 matrix nullBox = {};
2319 energyOutput.addDataAtEnergyStep(false, false, static_cast<double>(step),
2320 mdatoms->tmass, enerd, nullptr, nullptr, nullptr, nullBox,
2321 nullptr, nullptr, vir, pres, nullptr, mu_tot, constr);
2323 do_log = do_per_step(step, inputrec->nstlog);
2324 do_ene = do_per_step(step, inputrec->nstenergy);
2326 /* Prepare IMD energy record, if bIMD is TRUE. */
2327 IMD_fill_energy_record(inputrec->bIMD, inputrec->imd, enerd, step, TRUE);
2331 print_ebin_header(fplog, step, step);
2333 energyOutput.printStepToEnergyFile(mdoutf_get_fp_ene(outf), do_ene, FALSE, FALSE,
2334 do_log ? fplog : nullptr, step, step, eprNORMAL,
2335 fcd, &(top_global->groups), &(inputrec->opts), nullptr);
2338 /* Send x and E to IMD client, if bIMD is TRUE. */
2339 if (do_IMD(inputrec->bIMD, step, cr, TRUE, state_global->box, state_global->x.rvec_array(), inputrec, 0, wcycle) && MASTER(cr))
2341 IMD_send_positions(inputrec->imd);
2344 // Reset stepsize in we are doing more iterations
2345 stepsize = 1.0/ems.fnorm;
2347 /* Stop when the maximum force lies below tolerance.
2348 * If we have reached machine precision, converged is already set to true.
2350 converged = converged || (ems.fmax < inputrec->em_tol);
2352 } /* End of the loop */
2354 /* IMD cleanup, if bIMD is TRUE. */
2355 IMD_finalize(inputrec->bIMD, inputrec->imd);
2359 step--; /* we never took that last step in this case */
2362 if (ems.fmax > inputrec->em_tol)
2366 warn_step(fplog, inputrec->em_tol, ems.fmax,
2367 step-1 == number_steps, FALSE);
2372 /* If we printed energy and/or logfile last step (which was the last step)
2373 * we don't have to do it again, but otherwise print the final values.
2375 if (!do_log) /* Write final value to log since we didn't do anythin last step */
2377 print_ebin_header(fplog, step, step);
2379 if (!do_ene || !do_log) /* Write final energy file entries */
2381 energyOutput.printStepToEnergyFile(mdoutf_get_fp_ene(outf), !do_ene, FALSE, FALSE,
2382 !do_log ? fplog : nullptr, step, step, eprNORMAL,
2383 fcd, &(top_global->groups), &(inputrec->opts), nullptr);
2386 /* Print some stuff... */
2389 fprintf(stderr, "\nwriting lowest energy coordinates.\n");
2393 * For accurate normal mode calculation it is imperative that we
2394 * store the last conformation into the full precision binary trajectory.
2396 * However, we should only do it if we did NOT already write this step
2397 * above (which we did if do_x or do_f was true).
2399 do_x = !do_per_step(step, inputrec->nstxout);
2400 do_f = !do_per_step(step, inputrec->nstfout);
2401 write_em_traj(fplog, cr, outf, do_x, do_f, ftp2fn(efSTO, nfile, fnm),
2402 top_global, inputrec, step,
2403 &ems, state_global, observablesHistory);
2407 double sqrtNumAtoms = sqrt(static_cast<double>(state_global->natoms));
2408 print_converged(stderr, LBFGS, inputrec->em_tol, step, converged,
2409 number_steps, &ems, sqrtNumAtoms);
2410 print_converged(fplog, LBFGS, inputrec->em_tol, step, converged,
2411 number_steps, &ems, sqrtNumAtoms);
2413 fprintf(fplog, "\nPerformed %d energy evaluations in total.\n", neval);
2416 finish_em(cr, outf, walltime_accounting, wcycle);
2418 /* To print the actual number of steps we needed somewhere */
2419 walltime_accounting_set_nsteps_done(walltime_accounting, step);
2423 Integrator::do_steep()
2425 const char *SD = "Steepest Descents";
2427 gmx_enerdata_t *enerd;
2428 gmx_global_stat_t gstat;
2432 gmx_bool bDone, bAbort, do_x, do_f;
2437 int steps_accepted = 0;
2438 auto mdatoms = mdAtoms->mdatoms();
2440 GMX_LOG(mdlog.info).asParagraph().
2441 appendText("Note that activating steepest-descent energy minimization via the "
2442 "integrator .mdp option and the command gmx mdrun may "
2443 "be available in a different form in a future version of GROMACS, "
2444 "e.g. gmx minimize and an .mdp option.");
2446 /* Create 2 states on the stack and extract pointers that we will swap */
2447 em_state_t s0 {}, s1 {};
2448 em_state_t *s_min = &s0;
2449 em_state_t *s_try = &s1;
2451 /* Init em and store the local state in s_try */
2452 init_em(fplog, mdlog, SD, cr, ms, inputrec, mdrunOptions,
2453 state_global, top_global, s_try, &top,
2454 nrnb, fr, &graph, mdAtoms, &gstat,
2455 vsite, constr, nullptr,
2457 gmx_mdoutf *outf = init_mdoutf(fplog, nfile, fnm, mdrunOptions, cr, outputProvider, inputrec, top_global, nullptr, wcycle);
2459 init_enerdata(top_global->groups.grps[egcENER].nr, inputrec->fepvals->n_lambda, enerd);
2460 gmx::EnergyOutput energyOutput;
2461 energyOutput.prepare(mdoutf_get_fp_ene(outf), top_global, inputrec, nullptr);
2463 /* Print to log file */
2464 print_em_start(fplog, cr, walltime_accounting, wcycle, SD);
2466 /* Set variables for stepsize (in nm). This is the largest
2467 * step that we are going to make in any direction.
2469 ustep = inputrec->em_stepsize;
2472 /* Max number of steps */
2473 nsteps = inputrec->nsteps;
2477 /* Print to the screen */
2478 sp_header(stderr, SD, inputrec->em_tol, nsteps);
2482 sp_header(fplog, SD, inputrec->em_tol, nsteps);
2484 EnergyEvaluator energyEvaluator {
2485 fplog, mdlog, cr, ms,
2487 inputrec, nrnb, wcycle, gstat,
2488 vsite, constr, fcd, graph,
2489 mdAtoms, fr, ppForceWorkload, enerd
2492 /**** HERE STARTS THE LOOP ****
2493 * count is the counter for the number of steps
2494 * bDone will be TRUE when the minimization has converged
2495 * bAbort will be TRUE when nsteps steps have been performed or when
2496 * the stepsize becomes smaller than is reasonable for machine precision
2501 while (!bDone && !bAbort)
2503 bAbort = (nsteps >= 0) && (count == nsteps);
2505 /* set new coordinates, except for first step */
2506 bool validStep = true;
2510 do_em_step(cr, inputrec, mdatoms,
2511 s_min, stepsize, &s_min->f, s_try,
2517 energyEvaluator.run(s_try, mu_tot, vir, pres, count, count == 0);
2521 // Signal constraint error during stepping with energy=inf
2522 s_try->epot = std::numeric_limits<real>::infinity();
2527 print_ebin_header(fplog, count, count);
2532 s_min->epot = s_try->epot;
2535 /* Print it if necessary */
2538 if (mdrunOptions.verbose)
2540 fprintf(stderr, "Step=%5d, Dmax= %6.1e nm, Epot= %12.5e Fmax= %11.5e, atom= %d%c",
2541 count, ustep, s_try->epot, s_try->fmax, s_try->a_fmax+1,
2542 ( (count == 0) || (s_try->epot < s_min->epot) ) ? '\n' : '\r');
2546 if ( (count == 0) || (s_try->epot < s_min->epot) )
2548 /* Store the new (lower) energies */
2549 matrix nullBox = {};
2550 energyOutput.addDataAtEnergyStep(false, false, static_cast<double>(count),
2551 mdatoms->tmass, enerd, nullptr, nullptr, nullptr, nullBox,
2552 nullptr, nullptr, vir, pres, nullptr, mu_tot, constr);
2554 /* Prepare IMD energy record, if bIMD is TRUE. */
2555 IMD_fill_energy_record(inputrec->bIMD, inputrec->imd, enerd, count, TRUE);
2557 const bool do_dr = do_per_step(steps_accepted, inputrec->nstdisreout);
2558 const bool do_or = do_per_step(steps_accepted, inputrec->nstorireout);
2559 energyOutput.printStepToEnergyFile(mdoutf_get_fp_ene(outf), TRUE,
2561 fplog, count, count, eprNORMAL,
2562 fcd, &(top_global->groups),
2563 &(inputrec->opts), nullptr);
2568 /* Now if the new energy is smaller than the previous...
2569 * or if this is the first step!
2570 * or if we did random steps!
2573 if ( (count == 0) || (s_try->epot < s_min->epot) )
2577 /* Test whether the convergence criterion is met... */
2578 bDone = (s_try->fmax < inputrec->em_tol);
2580 /* Copy the arrays for force, positions and energy */
2581 /* The 'Min' array always holds the coords and forces of the minimal
2583 swap_em_state(&s_min, &s_try);
2589 /* Write to trn, if necessary */
2590 do_x = do_per_step(steps_accepted, inputrec->nstxout);
2591 do_f = do_per_step(steps_accepted, inputrec->nstfout);
2592 write_em_traj(fplog, cr, outf, do_x, do_f, nullptr,
2593 top_global, inputrec, count,
2594 s_min, state_global, observablesHistory);
2598 /* If energy is not smaller make the step smaller... */
2601 if (DOMAINDECOMP(cr) && s_min->s.ddp_count != cr->dd->ddp_count)
2603 /* Reload the old state */
2604 em_dd_partition_system(fplog, mdlog, count, cr, top_global, inputrec,
2605 s_min, &top, mdAtoms, fr, vsite, constr,
2610 /* Determine new step */
2611 stepsize = ustep/s_min->fmax;
2613 /* Check if stepsize is too small, with 1 nm as a characteristic length */
2615 if (count == nsteps || ustep < 1e-12)
2617 if (count == nsteps || ustep < 1e-6)
2622 warn_step(fplog, inputrec->em_tol, s_min->fmax,
2623 count == nsteps, constr != nullptr);
2628 /* Send IMD energies and positions, if bIMD is TRUE. */
2629 if (do_IMD(inputrec->bIMD, count, cr, TRUE, state_global->box,
2630 MASTER(cr) ? state_global->x.rvec_array() : nullptr,
2631 inputrec, 0, wcycle) &&
2634 IMD_send_positions(inputrec->imd);
2638 } /* End of the loop */
2640 /* IMD cleanup, if bIMD is TRUE. */
2641 IMD_finalize(inputrec->bIMD, inputrec->imd);
2643 /* Print some data... */
2646 fprintf(stderr, "\nwriting lowest energy coordinates.\n");
2648 write_em_traj(fplog, cr, outf, TRUE, inputrec->nstfout != 0, ftp2fn(efSTO, nfile, fnm),
2649 top_global, inputrec, count,
2650 s_min, state_global, observablesHistory);
2654 double sqrtNumAtoms = sqrt(static_cast<double>(state_global->natoms));
2656 print_converged(stderr, SD, inputrec->em_tol, count, bDone, nsteps,
2657 s_min, sqrtNumAtoms);
2658 print_converged(fplog, SD, inputrec->em_tol, count, bDone, nsteps,
2659 s_min, sqrtNumAtoms);
2662 finish_em(cr, outf, walltime_accounting, wcycle);
2664 /* To print the actual number of steps we needed somewhere */
2665 inputrec->nsteps = count;
2667 walltime_accounting_set_nsteps_done(walltime_accounting, count);
2673 const char *NM = "Normal Mode Analysis";
2676 gmx_enerdata_t *enerd;
2677 gmx_global_stat_t gstat;
2682 gmx_bool bSparse; /* use sparse matrix storage format */
2684 gmx_sparsematrix_t * sparse_matrix = nullptr;
2685 real * full_matrix = nullptr;
2687 /* added with respect to mdrun */
2689 real der_range = 10.0*std::sqrt(GMX_REAL_EPS);
2691 bool bIsMaster = MASTER(cr);
2692 auto mdatoms = mdAtoms->mdatoms();
2694 GMX_LOG(mdlog.info).asParagraph().
2695 appendText("Note that activating normal-mode analysis via the integrator "
2696 ".mdp option and the command gmx mdrun may "
2697 "be available in a different form in a future version of GROMACS, "
2698 "e.g. gmx normal-modes.");
2700 if (constr != nullptr)
2702 gmx_fatal(FARGS, "Constraints present with Normal Mode Analysis, this combination is not supported");
2705 gmx_shellfc_t *shellfc;
2707 em_state_t state_work {};
2709 /* Init em and store the local state in state_minimum */
2710 init_em(fplog, mdlog, NM, cr, ms, inputrec, mdrunOptions,
2711 state_global, top_global, &state_work, &top,
2712 nrnb, fr, &graph, mdAtoms, &gstat,
2713 vsite, constr, &shellfc,
2715 gmx_mdoutf *outf = init_mdoutf(fplog, nfile, fnm, mdrunOptions, cr, outputProvider, inputrec, top_global, nullptr, wcycle);
2717 init_enerdata(top_global->groups.grps[egcENER].nr, inputrec->fepvals->n_lambda, enerd);
2719 std::vector<int> atom_index = get_atom_index(top_global);
2720 std::vector<gmx::RVec> fneg(atom_index.size(), {0, 0, 0});
2721 snew(dfdx, atom_index.size());
2727 "NOTE: This version of GROMACS has been compiled in single precision,\n"
2728 " which MIGHT not be accurate enough for normal mode analysis.\n"
2729 " GROMACS now uses sparse matrix storage, so the memory requirements\n"
2730 " are fairly modest even if you recompile in double precision.\n\n");
2734 /* Check if we can/should use sparse storage format.
2736 * Sparse format is only useful when the Hessian itself is sparse, which it
2737 * will be when we use a cutoff.
2738 * For small systems (n<1000) it is easier to always use full matrix format, though.
2740 if (EEL_FULL(fr->ic->eeltype) || fr->rlist == 0.0)
2742 GMX_LOG(mdlog.warning).appendText("Non-cutoff electrostatics used, forcing full Hessian format.");
2745 else if (atom_index.size() < 1000)
2747 GMX_LOG(mdlog.warning).appendTextFormatted("Small system size (N=%zu), using full Hessian format.",
2753 GMX_LOG(mdlog.warning).appendText("Using compressed symmetric sparse Hessian format.");
2757 /* Number of dimensions, based on real atoms, that is not vsites or shell */
2758 sz = DIM*atom_index.size();
2760 fprintf(stderr, "Allocating Hessian memory...\n\n");
2764 sparse_matrix = gmx_sparsematrix_init(sz);
2765 sparse_matrix->compressed_symmetric = TRUE;
2769 snew(full_matrix, sz*sz);
2775 /* Write start time and temperature */
2776 print_em_start(fplog, cr, walltime_accounting, wcycle, NM);
2778 /* fudge nr of steps to nr of atoms */
2779 inputrec->nsteps = atom_index.size()*2;
2783 fprintf(stderr, "starting normal mode calculation '%s'\n%" PRId64 " steps.\n\n",
2784 *(top_global->name), inputrec->nsteps);
2787 nnodes = cr->nnodes;
2789 /* Make evaluate_energy do a single node force calculation */
2791 EnergyEvaluator energyEvaluator {
2792 fplog, mdlog, cr, ms,
2794 inputrec, nrnb, wcycle, gstat,
2795 vsite, constr, fcd, graph,
2796 mdAtoms, fr, ppForceWorkload, enerd
2798 energyEvaluator.run(&state_work, mu_tot, vir, pres, -1, TRUE);
2799 cr->nnodes = nnodes;
2801 /* if forces are not small, warn user */
2802 get_state_f_norm_max(cr, &(inputrec->opts), mdatoms, &state_work);
2804 GMX_LOG(mdlog.warning).appendTextFormatted("Maximum force:%12.5e", state_work.fmax);
2805 if (state_work.fmax > 1.0e-3)
2807 GMX_LOG(mdlog.warning).appendText(
2808 "The force is probably not small enough to "
2809 "ensure that you are at a minimum.\n"
2810 "Be aware that negative eigenvalues may occur\n"
2811 "when the resulting matrix is diagonalized.");
2814 /***********************************************************
2816 * Loop over all pairs in matrix
2818 * do_force called twice. Once with positive and
2819 * once with negative displacement
2821 ************************************************************/
2823 /* Steps are divided one by one over the nodes */
2825 auto state_work_x = makeArrayRef(state_work.s.x);
2826 auto state_work_f = makeArrayRef(state_work.f);
2827 for (unsigned int aid = cr->nodeid; aid < atom_index.size(); aid += nnodes)
2829 size_t atom = atom_index[aid];
2830 for (size_t d = 0; d < DIM; d++)
2833 int force_flags = GMX_FORCE_STATECHANGED | GMX_FORCE_ALLFORCES;
2836 x_min = state_work_x[atom][d];
2838 for (unsigned int dx = 0; (dx < 2); dx++)
2842 state_work_x[atom][d] = x_min - der_range;
2846 state_work_x[atom][d] = x_min + der_range;
2849 /* Make evaluate_energy do a single node force calculation */
2853 /* Now is the time to relax the shells */
2854 relax_shell_flexcon(fplog,
2857 mdrunOptions.verbose,
2868 state_work.f.arrayRefWithPadding(),
2874 &top_global->groups,
2881 DDBalanceRegionHandler(nullptr));
2887 energyEvaluator.run(&state_work, mu_tot, vir, pres, aid*2+dx, FALSE);
2890 cr->nnodes = nnodes;
2894 std::copy(state_work_f.begin(), state_work_f.begin()+atom_index.size(), fneg.begin());
2898 /* x is restored to original */
2899 state_work_x[atom][d] = x_min;
2901 for (size_t j = 0; j < atom_index.size(); j++)
2903 for (size_t k = 0; (k < DIM); k++)
2906 -(state_work_f[atom_index[j]][k] - fneg[j][k])/(2*der_range);
2913 #define mpi_type GMX_MPI_REAL
2914 MPI_Send(dfdx[0], atom_index.size()*DIM, mpi_type, MASTER(cr),
2915 cr->nodeid, cr->mpi_comm_mygroup);
2920 for (node = 0; (node < nnodes && aid+node < atom_index.size()); node++)
2926 MPI_Recv(dfdx[0], atom_index.size()*DIM, mpi_type, node, node,
2927 cr->mpi_comm_mygroup, &stat);
2932 row = (aid + node)*DIM + d;
2934 for (size_t j = 0; j < atom_index.size(); j++)
2936 for (size_t k = 0; k < DIM; k++)
2942 if (col >= row && dfdx[j][k] != 0.0)
2944 gmx_sparsematrix_increment_value(sparse_matrix,
2945 row, col, dfdx[j][k]);
2950 full_matrix[row*sz+col] = dfdx[j][k];
2957 if (mdrunOptions.verbose && fplog)
2962 /* write progress */
2963 if (bIsMaster && mdrunOptions.verbose)
2965 fprintf(stderr, "\rFinished step %d out of %td",
2966 std::min<int>(atom+nnodes, atom_index.size()),
2974 fprintf(stderr, "\n\nWriting Hessian...\n");
2975 gmx_mtxio_write(ftp2fn(efMTX, nfile, fnm), sz, sz, full_matrix, sparse_matrix);
2978 finish_em(cr, outf, walltime_accounting, wcycle);
2980 walltime_accounting_set_nsteps_done(walltime_accounting, atom_index.size()*2);