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) 2011-2019,2020,2021, 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 Implements the integrator for normal molecular dynamics simulations
41 * \author David van der Spoel <david.vanderspoel@icm.uu.se>
42 * \ingroup module_mdrun
55 #include "gromacs/applied_forces/awh/awh.h"
56 #include "gromacs/applied_forces/awh/read_params.h"
57 #include "gromacs/commandline/filenm.h"
58 #include "gromacs/domdec/collect.h"
59 #include "gromacs/domdec/dlbtiming.h"
60 #include "gromacs/domdec/domdec.h"
61 #include "gromacs/domdec/domdec_network.h"
62 #include "gromacs/domdec/domdec_struct.h"
63 #include "gromacs/domdec/gpuhaloexchange.h"
64 #include "gromacs/domdec/localtopologychecker.h"
65 #include "gromacs/domdec/mdsetup.h"
66 #include "gromacs/domdec/partition.h"
67 #include "gromacs/essentialdynamics/edsam.h"
68 #include "gromacs/ewald/pme_load_balancing.h"
69 #include "gromacs/ewald/pme_pp.h"
70 #include "gromacs/fileio/trxio.h"
71 #include "gromacs/gmxlib/network.h"
72 #include "gromacs/gmxlib/nrnb.h"
73 #include "gromacs/gpu_utils/device_stream_manager.h"
74 #include "gromacs/gpu_utils/gpu_utils.h"
75 #include "gromacs/imd/imd.h"
76 #include "gromacs/listed_forces/listed_forces.h"
77 #include "gromacs/math/functions.h"
78 #include "gromacs/math/invertmatrix.h"
79 #include "gromacs/math/vec.h"
80 #include "gromacs/math/vectypes.h"
81 #include "gromacs/mdlib/checkpointhandler.h"
82 #include "gromacs/mdlib/compute_io.h"
83 #include "gromacs/mdlib/constr.h"
84 #include "gromacs/mdlib/coupling.h"
85 #include "gromacs/mdlib/ebin.h"
86 #include "gromacs/mdlib/enerdata_utils.h"
87 #include "gromacs/mdlib/energyoutput.h"
88 #include "gromacs/mdlib/expanded.h"
89 #include "gromacs/mdlib/force.h"
90 #include "gromacs/mdlib/force_flags.h"
91 #include "gromacs/mdlib/forcerec.h"
92 #include "gromacs/mdlib/freeenergyparameters.h"
93 #include "gromacs/mdlib/md_support.h"
94 #include "gromacs/mdlib/mdatoms.h"
95 #include "gromacs/mdlib/mdoutf.h"
96 #include "gromacs/mdlib/membed.h"
97 #include "gromacs/mdlib/resethandler.h"
98 #include "gromacs/mdlib/sighandler.h"
99 #include "gromacs/mdlib/simulationsignal.h"
100 #include "gromacs/mdlib/stat.h"
101 #include "gromacs/mdlib/stophandler.h"
102 #include "gromacs/mdlib/tgroup.h"
103 #include "gromacs/mdlib/trajectory_writing.h"
104 #include "gromacs/mdlib/update.h"
105 #include "gromacs/mdlib/update_constrain_gpu.h"
106 #include "gromacs/mdlib/update_vv.h"
107 #include "gromacs/mdlib/vcm.h"
108 #include "gromacs/mdlib/vsite.h"
109 #include "gromacs/mdrunutility/freeenergy.h"
110 #include "gromacs/mdrunutility/handlerestart.h"
111 #include "gromacs/mdrunutility/multisim.h"
112 #include "gromacs/mdrunutility/printtime.h"
113 #include "gromacs/mdtypes/awh_history.h"
114 #include "gromacs/mdtypes/awh_params.h"
115 #include "gromacs/mdtypes/commrec.h"
116 #include "gromacs/mdtypes/df_history.h"
117 #include "gromacs/mdtypes/energyhistory.h"
118 #include "gromacs/mdtypes/fcdata.h"
119 #include "gromacs/mdtypes/forcebuffers.h"
120 #include "gromacs/mdtypes/forcerec.h"
121 #include "gromacs/mdtypes/group.h"
122 #include "gromacs/mdtypes/inputrec.h"
123 #include "gromacs/mdtypes/interaction_const.h"
124 #include "gromacs/mdtypes/md_enums.h"
125 #include "gromacs/mdtypes/mdatom.h"
126 #include "gromacs/mdtypes/mdrunoptions.h"
127 #include "gromacs/mdtypes/multipletimestepping.h"
128 #include "gromacs/mdtypes/observableshistory.h"
129 #include "gromacs/mdtypes/observablesreducer.h"
130 #include "gromacs/mdtypes/pullhistory.h"
131 #include "gromacs/mdtypes/simulation_workload.h"
132 #include "gromacs/mdtypes/state.h"
133 #include "gromacs/mdtypes/state_propagator_data_gpu.h"
134 #include "gromacs/modularsimulator/energydata.h"
135 #include "gromacs/nbnxm/gpu_data_mgmt.h"
136 #include "gromacs/nbnxm/nbnxm.h"
137 #include "gromacs/pbcutil/pbc.h"
138 #include "gromacs/pulling/output.h"
139 #include "gromacs/pulling/pull.h"
140 #include "gromacs/swap/swapcoords.h"
141 #include "gromacs/timing/wallcycle.h"
142 #include "gromacs/timing/walltime_accounting.h"
143 #include "gromacs/topology/atoms.h"
144 #include "gromacs/topology/idef.h"
145 #include "gromacs/topology/mtop_util.h"
146 #include "gromacs/topology/topology.h"
147 #include "gromacs/trajectory/trajectoryframe.h"
148 #include "gromacs/utility/basedefinitions.h"
149 #include "gromacs/utility/cstringutil.h"
150 #include "gromacs/utility/fatalerror.h"
151 #include "gromacs/utility/logger.h"
152 #include "gromacs/utility/real.h"
153 #include "gromacs/utility/smalloc.h"
155 #include "legacysimulator.h"
156 #include "replicaexchange.h"
159 using gmx::SimulationSignaller;
161 void gmx::LegacySimulator::do_md()
163 // TODO Historically, the EM and MD "integrators" used different
164 // names for the t_inputrec *parameter, but these must have the
165 // same name, now that it's a member of a struct. We use this ir
166 // alias to avoid a large ripple of nearly useless changes.
167 // t_inputrec is being replaced by IMdpOptionsProvider, so this
168 // will go away eventually.
169 const t_inputrec* ir = inputrec;
171 double t, t0 = ir->init_t;
172 gmx_bool bGStatEveryStep, bGStat, bCalcVir, bCalcEnerStep, bCalcEner;
173 gmx_bool bNS = FALSE, bNStList, bStopCM, bFirstStep, bInitStep, bLastStep = FALSE;
174 gmx_bool bDoExpanded = FALSE;
175 gmx_bool do_ene, do_log, do_verbose;
176 gmx_bool bMasterState;
177 unsigned int force_flags;
178 tensor force_vir = { { 0 } }, shake_vir = { { 0 } }, total_vir = { { 0 } }, pres = { { 0 } };
181 matrix pressureCouplingMu, M;
182 gmx_repl_ex_t repl_ex = nullptr;
183 gmx_global_stat_t gstat;
184 gmx_shellfc_t* shellfc;
185 gmx_bool bSumEkinhOld, bDoReplEx, bExchanged, bNeedRepartition;
188 std::vector<RVec> cbuf;
193 real saved_conserved_quantity = 0;
196 char sbuf[STEPSTRSIZE], sbuf2[STEPSTRSIZE];
198 /* PME load balancing data for GPU kernels */
199 gmx_bool bPMETune = FALSE;
200 gmx_bool bPMETunePrinting = FALSE;
202 bool bInteractiveMDstep = false;
204 SimulationSignals signals;
205 // Most global communnication stages don't propagate mdrun
206 // signals, and will use this object to achieve that.
207 SimulationSignaller nullSignaller(nullptr, nullptr, nullptr, false, false);
209 if (!mdrunOptions.writeConfout)
211 // This is on by default, and the main known use case for
212 // turning it off is for convenience in benchmarking, which is
213 // something that should not show up in the general user
218 "The -noconfout functionality is deprecated, and may be removed in a "
222 /* md-vv uses averaged full step velocities for T-control
223 md-vv-avek uses averaged half step velocities for T-control (but full step ekin for P control)
224 md uses averaged half step kinetic energies to determine temperature unless defined otherwise by GMX_EKIN_AVE_VEL; */
225 bTrotter = (EI_VV(ir->eI)
226 && (inputrecNptTrotter(ir) || inputrecNphTrotter(ir) || inputrecNvtTrotter(ir)));
228 const bool bRerunMD = false;
230 int nstglobalcomm = computeGlobalCommunicationPeriod(mdlog, ir, cr);
231 bGStatEveryStep = (nstglobalcomm == 1);
233 const SimulationGroups* groups = &top_global.groups;
235 std::unique_ptr<EssentialDynamics> ed = nullptr;
236 if (opt2bSet("-ei", nfile, fnm))
238 /* Initialize essential dynamics sampling */
239 ed = init_edsam(mdlog,
240 opt2fn_null("-ei", nfile, fnm),
241 opt2fn("-eo", nfile, fnm),
251 else if (observablesHistory->edsamHistory)
254 "The checkpoint is from a run with essential dynamics sampling, "
255 "but the current run did not specify the -ei option. "
256 "Either specify the -ei option to mdrun, or do not use this checkpoint file.");
259 int* fep_state = MASTER(cr) ? &state_global->fep_state : nullptr;
260 gmx::ArrayRef<real> lambda = MASTER(cr) ? state_global->lambda : gmx::ArrayRef<real>();
261 initialize_lambdas(fplog,
265 ir->simtempvals->temperatures,
266 gmx::arrayRefFromArray(ir->opts.ref_t, ir->opts.ngtc),
270 Update upd(*ir, deform);
271 bool doSimulatedAnnealing = false;
273 // TODO: Avoid changing inputrec (#3854)
274 // Simulated annealing updates the reference temperature.
275 auto* nonConstInputrec = const_cast<t_inputrec*>(inputrec);
276 doSimulatedAnnealing = initSimulatedAnnealing(nonConstInputrec, &upd);
278 const bool useReplicaExchange = (replExParams.exchangeInterval > 0);
280 t_fcdata& fcdata = *fr->fcdata;
282 bool simulationsShareState = false;
283 int nstSignalComm = nstglobalcomm;
285 // TODO This implementation of ensemble orientation restraints is nasty because
286 // a user can't just do multi-sim with single-sim orientation restraints.
287 bool usingEnsembleRestraints = (fcdata.disres->nsystems > 1) || ((ms != nullptr) && fcdata.orires);
288 bool awhUsesMultiSim = (ir->bDoAwh && ir->awhParams->shareBiasMultisim() && (ms != nullptr));
290 // Replica exchange, ensemble restraints and AWH need all
291 // simulations to remain synchronized, so they need
292 // checkpoints and stop conditions to act on the same step, so
293 // the propagation of such signals must take place between
294 // simulations, not just within simulations.
295 // TODO: Make algorithm initializers set these flags.
296 simulationsShareState = useReplicaExchange || usingEnsembleRestraints || awhUsesMultiSim;
298 if (simulationsShareState)
300 // Inter-simulation signal communication does not need to happen
301 // often, so we use a minimum of 200 steps to reduce overhead.
302 const int c_minimumInterSimulationSignallingInterval = 200;
303 nstSignalComm = ((c_minimumInterSimulationSignallingInterval + nstglobalcomm - 1) / nstglobalcomm)
308 if (startingBehavior != StartingBehavior::RestartWithAppending)
310 pleaseCiteCouplingAlgorithms(fplog, *ir);
312 gmx_mdoutf* outf = init_mdoutf(fplog,
324 simulationsShareState,
326 gmx::EnergyOutput energyOutput(mdoutf_get_fp_ene(outf),
330 mdoutf_get_fp_dhdl(outf),
333 simulationsShareState,
336 gstat = global_stat_init(ir);
338 const auto& simulationWork = runScheduleWork->simulationWork;
339 const bool useGpuForPme = simulationWork.useGpuPme;
340 const bool useGpuForNonbonded = simulationWork.useGpuNonbonded;
341 const bool useGpuForBufferOps = simulationWork.useGpuBufferOps;
342 const bool useGpuForUpdate = simulationWork.useGpuUpdate;
344 /* Check for polarizable models and flexible constraints */
345 shellfc = init_shell_flexcon(fplog,
347 constr ? constr->numFlexibleConstraints() : 0,
349 haveDDAtomOrdering(*cr),
353 double io = compute_io(ir, top_global.natoms, *groups, energyOutput.numEnergyTerms(), 1);
354 if ((io > 2000) && MASTER(cr))
356 fprintf(stderr, "\nWARNING: This run will generate roughly %.0f Mb of data\n\n", io);
360 ObservablesReducer observablesReducer = observablesReducerBuilder->build();
362 ForceBuffers f(simulationWork.useMts,
363 ((useGpuForNonbonded && useGpuForBufferOps) || useGpuForUpdate)
364 ? PinningPolicy::PinnedIfSupported
365 : PinningPolicy::CannotBePinned);
366 const t_mdatoms* md = mdAtoms->mdatoms();
367 if (haveDDAtomOrdering(*cr))
369 // Local state only becomes valid now.
370 dd_init_local_state(*cr->dd, state_global, state);
372 /* Distribute the charge groups over the nodes from the master node */
373 dd_partition_system(fplog,
394 upd.updateAfterPartition(state->natoms,
395 md->cFREEZE ? gmx::arrayRefFromArray(md->cFREEZE, md->nr)
396 : gmx::ArrayRef<const unsigned short>(),
397 md->cTC ? gmx::arrayRefFromArray(md->cTC, md->nr)
398 : gmx::ArrayRef<const unsigned short>());
399 fr->longRangeNonbondeds->updateAfterPartition(*md);
403 state_change_natoms(state_global, state_global->natoms);
405 /* Generate and initialize new topology */
406 mdAlgorithmsSetupAtomData(cr, *ir, top_global, top, fr, &f, mdAtoms, constr, vsite, shellfc);
408 upd.updateAfterPartition(state->natoms,
409 md->cFREEZE ? gmx::arrayRefFromArray(md->cFREEZE, md->nr)
410 : gmx::ArrayRef<const unsigned short>(),
411 md->cTC ? gmx::arrayRefFromArray(md->cTC, md->nr)
412 : gmx::ArrayRef<const unsigned short>());
413 fr->longRangeNonbondeds->updateAfterPartition(*md);
416 std::unique_ptr<UpdateConstrainGpu> integrator;
418 StatePropagatorDataGpu* stateGpu = fr->stateGpu;
420 // TODO: the assertions below should be handled by UpdateConstraintsBuilder.
423 GMX_RELEASE_ASSERT(!haveDDAtomOrdering(*cr) || ddUsesUpdateGroups(*cr->dd)
424 || constr == nullptr || constr->numConstraintsTotal() == 0,
425 "Constraints in domain decomposition are only supported with update "
426 "groups if using GPU update.\n");
427 GMX_RELEASE_ASSERT(ir->eConstrAlg != ConstraintAlgorithm::Shake || constr == nullptr
428 || constr->numConstraintsTotal() == 0,
429 "SHAKE is not supported with GPU update.");
430 GMX_RELEASE_ASSERT(useGpuForPme || (useGpuForNonbonded && simulationWork.useGpuBufferOps),
431 "Either PME or short-ranged non-bonded interaction tasks must run on "
432 "the GPU to use GPU update.\n");
433 GMX_RELEASE_ASSERT(ir->eI == IntegrationAlgorithm::MD,
434 "Only the md integrator is supported with the GPU update.\n");
436 ir->etc != TemperatureCoupling::NoseHoover,
437 "Nose-Hoover temperature coupling is not supported with the GPU update.\n");
439 ir->epc == PressureCoupling::No || ir->epc == PressureCoupling::ParrinelloRahman
440 || ir->epc == PressureCoupling::Berendsen || ir->epc == PressureCoupling::CRescale,
441 "Only Parrinello-Rahman, Berendsen, and C-rescale pressure coupling are supported "
442 "with the GPU update.\n");
443 GMX_RELEASE_ASSERT(!md->haveVsites,
444 "Virtual sites are not supported with the GPU update.\n");
445 GMX_RELEASE_ASSERT(ed == nullptr,
446 "Essential dynamics is not supported with the GPU update.\n");
447 GMX_RELEASE_ASSERT(!ir->bPull || !pull_have_constraint(*ir->pull),
448 "Constraints pulling is not supported with the GPU update.\n");
449 GMX_RELEASE_ASSERT(fcdata.orires == nullptr,
450 "Orientation restraints are not supported with the GPU update.\n");
452 ir->efep == FreeEnergyPerturbationType::No
453 || (!haveFepPerturbedMasses(top_global) && !havePerturbedConstraints(top_global)),
454 "Free energy perturbation of masses and constraints are not supported with the GPU "
457 if (constr != nullptr && constr->numConstraintsTotal() > 0)
461 .appendText("Updating coordinates and applying constraints on the GPU.");
465 GMX_LOG(mdlog.info).asParagraph().appendText("Updating coordinates on the GPU.");
467 GMX_RELEASE_ASSERT(fr->deviceStreamManager != nullptr,
468 "Device stream manager should be initialized in order to use GPU "
469 "update-constraints.");
471 fr->deviceStreamManager->streamIsValid(gmx::DeviceStreamType::UpdateAndConstraints),
472 "Update stream should be initialized in order to use GPU "
473 "update-constraints.");
474 integrator = std::make_unique<UpdateConstrainGpu>(
478 fr->deviceStreamManager->context(),
479 fr->deviceStreamManager->stream(gmx::DeviceStreamType::UpdateAndConstraints),
482 stateGpu->setXUpdatedOnDeviceEvent(integrator->xUpdatedOnDeviceEvent());
484 integrator->setPbc(PbcType::Xyz, state->box);
487 if (useGpuForPme || (useGpuForNonbonded && useGpuForBufferOps) || useGpuForUpdate)
489 changePinningPolicy(&state->x, PinningPolicy::PinnedIfSupported);
493 changePinningPolicy(&state->v, PinningPolicy::PinnedIfSupported);
496 // NOTE: The global state is no longer used at this point.
497 // But state_global is still used as temporary storage space for writing
498 // the global state to file and potentially for replica exchange.
499 // (Global topology should persist.)
501 update_mdatoms(mdAtoms->mdatoms(), state->lambda[FreeEnergyPerturbationCouplingType::Mass]);
505 /* Check nstexpanded here, because the grompp check was broken */
506 if (ir->expandedvals->nstexpanded % ir->nstcalcenergy != 0)
509 "With expanded ensemble, nstexpanded should be a multiple of nstcalcenergy");
511 init_expanded_ensemble(startingBehavior != StartingBehavior::NewSimulation, ir, state->dfhist);
516 EnergyData::initializeEnergyHistory(startingBehavior, observablesHistory, &energyOutput);
519 preparePrevStepPullCom(ir,
521 gmx::arrayRefFromArray(md->massT, md->nr),
525 startingBehavior != StartingBehavior::NewSimulation);
527 // TODO: Remove this by converting AWH into a ForceProvider
528 auto awh = prepareAwhModule(fplog,
533 startingBehavior != StartingBehavior::NewSimulation,
535 opt2fn("-awh", nfile, fnm),
538 if (useReplicaExchange && MASTER(cr))
540 repl_ex = init_replica_exchange(fplog, ms, top_global.natoms, ir, replExParams);
542 /* PME tuning is only supported in the Verlet scheme, with PME for
543 * Coulomb. It is not supported with only LJ PME. */
544 bPMETune = (mdrunOptions.tunePme && EEL_PME(fr->ic->eeltype) && !mdrunOptions.reproducible
545 && ir->cutoff_scheme != CutoffScheme::Group);
547 pme_load_balancing_t* pme_loadbal = nullptr;
551 &pme_loadbal, cr, mdlog, *ir, state->box, *fr->ic, *fr->nbv, fr->pmedata, fr->nbv->useGpu());
554 if (!ir->bContinuation)
556 if (state->flags & enumValueToBitMask(StateEntry::V))
558 auto v = makeArrayRef(state->v);
559 /* Set the velocities of vsites, shells and frozen atoms to zero */
560 for (i = 0; i < md->homenr; i++)
562 if (md->ptype[i] == ParticleType::Shell)
566 else if (md->cFREEZE)
568 for (m = 0; m < DIM; m++)
570 if (ir->opts.nFreeze[md->cFREEZE[i]][m])
581 /* Constrain the initial coordinates and velocities */
582 do_constrain_first(fplog,
587 state->x.arrayRefWithPadding(),
588 state->v.arrayRefWithPadding(),
590 state->lambda[FreeEnergyPerturbationCouplingType::Bonded]);
594 const int nstfep = computeFepPeriod(*ir, replExParams);
596 /* Be REALLY careful about what flags you set here. You CANNOT assume
597 * this is the first step, since we might be restarting from a checkpoint,
598 * and in that case we should not do any modifications to the state.
600 bStopCM = (ir->comm_mode != ComRemovalAlgorithm::No && !ir->bContinuation);
602 // When restarting from a checkpoint, it can be appropriate to
603 // initialize ekind from quantities in the checkpoint. Otherwise,
604 // compute_globals must initialize ekind before the simulation
605 // starts/restarts. However, only the master rank knows what was
606 // found in the checkpoint file, so we have to communicate in
607 // order to coordinate the restart.
609 // TODO Consider removing this communication if/when checkpoint
610 // reading directly follows .tpr reading, because all ranks can
611 // agree on hasReadEkinState at that time.
612 bool hasReadEkinState = MASTER(cr) ? state_global->ekinstate.hasReadEkinState : false;
615 gmx_bcast(sizeof(hasReadEkinState), &hasReadEkinState, cr->mpi_comm_mygroup);
617 if (hasReadEkinState)
619 restore_ekinstate_from_state(cr, ekind, &state_global->ekinstate);
622 unsigned int cglo_flags =
623 (CGLO_TEMPERATURE | CGLO_GSTAT | (EI_VV(ir->eI) ? CGLO_PRESSURE : 0)
624 | (EI_VV(ir->eI) ? CGLO_CONSTRAINT : 0) | (hasReadEkinState ? CGLO_READEKIN : 0));
626 bSumEkinhOld = FALSE;
628 t_vcm vcm(top_global.groups, *ir);
629 reportComRemovalInfo(fplog, vcm);
631 int64_t step = ir->init_step;
632 int64_t step_rel = 0;
634 /* To minimize communication, compute_globals computes the COM velocity
635 * and the kinetic energy for the velocities without COM motion removed.
636 * Thus to get the kinetic energy without the COM contribution, we need
637 * to call compute_globals twice.
639 for (int cgloIteration = 0; cgloIteration < (bStopCM ? 2 : 1); cgloIteration++)
641 unsigned int cglo_flags_iteration = cglo_flags;
642 if (bStopCM && cgloIteration == 0)
644 cglo_flags_iteration |= CGLO_STOPCM;
645 cglo_flags_iteration &= ~CGLO_TEMPERATURE;
647 compute_globals(gstat,
652 makeConstArrayRef(state->x),
653 makeConstArrayRef(state->v),
667 cglo_flags_iteration,
669 &observablesReducer);
670 // Clean up after pre-step use of compute_globals()
671 observablesReducer.markAsReadyToReduce();
673 if (cglo_flags_iteration & CGLO_STOPCM)
675 /* At initialization, do not pass x with acceleration-correction mode
676 * to avoid (incorrect) correction of the initial coordinates.
678 auto x = (vcm.mode == ComRemovalAlgorithm::LinearAccelerationCorrection)
680 : makeArrayRef(state->x);
681 process_and_stopcm_grp(fplog, &vcm, *md, x, makeArrayRef(state->v));
682 inc_nrnb(nrnb, eNR_STOPCM, md->homenr);
685 if (ir->eI == IntegrationAlgorithm::VVAK)
687 /* a second call to get the half step temperature initialized as well */
688 /* we do the same call as above, but turn the pressure off -- internally to
689 compute_globals, this is recognized as a velocity verlet half-step
690 kinetic energy calculation. This minimized excess variables, but
691 perhaps loses some logic?*/
693 compute_globals(gstat,
698 makeConstArrayRef(state->x),
699 makeConstArrayRef(state->v),
713 cglo_flags & ~CGLO_PRESSURE,
715 &observablesReducer);
716 // Clean up after pre-step use of compute_globals()
717 observablesReducer.markAsReadyToReduce();
720 /* Calculate the initial half step temperature, and save the ekinh_old */
721 if (startingBehavior == StartingBehavior::NewSimulation)
723 for (i = 0; (i < ir->opts.ngtc); i++)
725 copy_mat(ekind->tcstat[i].ekinh, ekind->tcstat[i].ekinh_old);
729 /* need to make an initiation call to get the Trotter variables set, as well as other constants
730 for non-trotter temperature control */
731 auto trotter_seq = init_npt_vars(ir, state, &MassQ, bTrotter);
735 if (!ir->bContinuation)
737 if (constr && ir->eConstrAlg == ConstraintAlgorithm::Lincs)
740 "RMS relative constraint deviation after constraining: %.2e\n",
743 if (EI_STATE_VELOCITY(ir->eI))
745 real temp = enerd->term[F_TEMP];
746 if (ir->eI != IntegrationAlgorithm::VV)
748 /* Result of Ekin averaged over velocities of -half
749 * and +half step, while we only have -half step here.
753 fprintf(fplog, "Initial temperature: %g K\n", temp);
758 fprintf(stderr, "starting mdrun '%s'\n", *(top_global.name));
761 sprintf(tbuf, "%8.1f", (ir->init_step + ir->nsteps) * ir->delta_t);
765 sprintf(tbuf, "%s", "infinite");
767 if (ir->init_step > 0)
770 "%s steps, %s ps (continuing from step %s, %8.1f ps).\n",
771 gmx_step_str(ir->init_step + ir->nsteps, sbuf),
773 gmx_step_str(ir->init_step, sbuf2),
774 ir->init_step * ir->delta_t);
778 fprintf(stderr, "%s steps, %s ps.\n", gmx_step_str(ir->nsteps, sbuf), tbuf);
780 fprintf(fplog, "\n");
783 walltime_accounting_start_time(walltime_accounting);
784 wallcycle_start(wcycle, WallCycleCounter::Run);
785 print_start(fplog, cr, walltime_accounting, "mdrun");
787 /***********************************************************
791 ************************************************************/
794 /* Skip the first Nose-Hoover integration when we get the state from tpx */
795 bInitStep = startingBehavior == StartingBehavior::NewSimulation || EI_VV(ir->eI);
796 bSumEkinhOld = FALSE;
798 bNeedRepartition = FALSE;
800 auto stopHandler = stopHandlerBuilder->getStopHandlerMD(
801 compat::not_null<SimulationSignal*>(&signals[eglsSTOPCOND]),
802 simulationsShareState,
805 mdrunOptions.reproducible,
807 mdrunOptions.maximumHoursToRun,
812 walltime_accounting);
814 auto checkpointHandler = std::make_unique<CheckpointHandler>(
815 compat::make_not_null<SimulationSignal*>(&signals[eglsCHKPT]),
816 simulationsShareState,
819 mdrunOptions.writeConfout,
820 mdrunOptions.checkpointOptions.period);
822 const bool resetCountersIsLocal = true;
823 auto resetHandler = std::make_unique<ResetHandler>(
824 compat::make_not_null<SimulationSignal*>(&signals[eglsRESETCOUNTERS]),
825 !resetCountersIsLocal,
828 mdrunOptions.timingOptions.resetHalfway,
829 mdrunOptions.maximumHoursToRun,
832 walltime_accounting);
834 const DDBalanceRegionHandler ddBalanceRegionHandler(cr);
836 if (MASTER(cr) && isMultiSim(ms) && !useReplicaExchange)
838 logInitialMultisimStatus(ms, cr, mdlog, simulationsShareState, ir->nsteps, ir->init_step);
841 /* and stop now if we should */
842 bLastStep = (bLastStep || (ir->nsteps >= 0 && step_rel > ir->nsteps));
846 /* Determine if this is a neighbor search step */
847 bNStList = (ir->nstlist > 0 && step % ir->nstlist == 0);
849 if (bPMETune && bNStList)
851 // This has to be here because PME load balancing is called so early.
852 // TODO: Move to after all booleans are defined.
853 if (useGpuForUpdate && !bFirstStep)
855 stateGpu->copyCoordinatesFromGpu(state->x, AtomLocality::Local);
856 stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
858 /* PME grid + cut-off optimization with GPUs or PME nodes */
859 pme_loadbal_do(pme_loadbal,
861 (mdrunOptions.verbose && MASTER(cr)) ? stderr : nullptr,
872 simulationWork.useGpuPmePpCommunication);
875 wallcycle_start(wcycle, WallCycleCounter::Step);
877 bLastStep = (step_rel == ir->nsteps);
878 t = t0 + step * ir->delta_t;
880 // TODO Refactor this, so that nstfep does not need a default value of zero
881 if (ir->efep != FreeEnergyPerturbationType::No || ir->bSimTemp)
883 /* find and set the current lambdas */
884 state->lambda = currentLambdas(step, *(ir->fepvals), state->fep_state);
886 bDoExpanded = (do_per_step(step, ir->expandedvals->nstexpanded) && (ir->bExpanded)
890 bDoReplEx = (useReplicaExchange && (step > 0) && !bLastStep
891 && do_per_step(step, replExParams.exchangeInterval));
893 if (doSimulatedAnnealing)
895 // TODO: Avoid changing inputrec (#3854)
896 // Simulated annealing updates the reference temperature.
897 auto* nonConstInputrec = const_cast<t_inputrec*>(inputrec);
898 update_annealing_target_temp(nonConstInputrec, t, &upd);
901 /* Stop Center of Mass motion */
902 bStopCM = (ir->comm_mode != ComRemovalAlgorithm::No && do_per_step(step, ir->nstcomm));
904 /* Determine whether or not to do Neighbour Searching */
905 bNS = (bFirstStep || bNStList || bExchanged || bNeedRepartition);
907 /* Note that the stopHandler will cause termination at nstglobalcomm
908 * steps. Since this concides with nstcalcenergy, nsttcouple and/or
909 * nstpcouple steps, we have computed the half-step kinetic energy
910 * of the previous step and can always output energies at the last step.
912 bLastStep = bLastStep || stopHandler->stoppingAfterCurrentStep(bNS);
914 /* do_log triggers energy and virial calculation. Because this leads
915 * to different code paths, forces can be different. Thus for exact
916 * continuation we should avoid extra log output.
917 * Note that the || bLastStep can result in non-exact continuation
918 * beyond the last step. But we don't consider that to be an issue.
920 do_log = (do_per_step(step, ir->nstlog)
921 || (bFirstStep && startingBehavior == StartingBehavior::NewSimulation) || bLastStep);
922 do_verbose = mdrunOptions.verbose
923 && (step % mdrunOptions.verboseStepPrintInterval == 0 || bFirstStep || bLastStep);
925 // On search steps, when doing the update on the GPU, copy
926 // the coordinates and velocities to the host unless they are
927 // already there (ie on the first step and after replica
929 if (useGpuForUpdate && bNS && !bFirstStep && !bExchanged)
931 stateGpu->copyVelocitiesFromGpu(state->v, AtomLocality::Local);
932 stateGpu->copyCoordinatesFromGpu(state->x, AtomLocality::Local);
933 stateGpu->waitVelocitiesReadyOnHost(AtomLocality::Local);
934 stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
937 // We only need to calculate virtual velocities if we are writing them in the current step
938 const bool needVirtualVelocitiesThisStep =
940 && (do_per_step(step, ir->nstvout) || checkpointHandler->isCheckpointingStep());
942 if (vsite != nullptr)
944 // Virtual sites need to be updated before domain decomposition and forces are calculated
945 wallcycle_start(wcycle, WallCycleCounter::VsiteConstr);
946 // md-vv calculates virtual velocities once it has full-step real velocities
947 vsite->construct(state->x,
950 (!EI_VV(inputrec->eI) && needVirtualVelocitiesThisStep)
951 ? VSiteOperation::PositionsAndVelocities
952 : VSiteOperation::Positions);
953 wallcycle_stop(wcycle, WallCycleCounter::VsiteConstr);
956 if (bNS && !(bFirstStep && ir->bContinuation))
958 bMasterState = FALSE;
959 /* Correct the new box if it is too skewed */
960 if (inputrecDynamicBox(ir))
962 if (correct_box(fplog, step, state->box))
967 // If update is offloaded, and the box was changed either
968 // above or in a replica exchange on the previous step,
969 // the GPU Update object should be informed
970 if (useGpuForUpdate && (bMasterState || bExchanged))
972 integrator->setPbc(PbcType::Xyz, state->box);
974 if (haveDDAtomOrdering(*cr) && bMasterState)
976 dd_collect_state(cr->dd, state, state_global);
979 if (haveDDAtomOrdering(*cr))
981 /* Repartition the domain decomposition */
982 dd_partition_system(fplog,
1002 do_verbose && !bPMETunePrinting);
1003 upd.updateAfterPartition(state->natoms,
1004 md->cFREEZE ? gmx::arrayRefFromArray(md->cFREEZE, md->nr)
1005 : gmx::ArrayRef<const unsigned short>(),
1006 md->cTC ? gmx::arrayRefFromArray(md->cTC, md->nr)
1007 : gmx::ArrayRef<const unsigned short>());
1008 fr->longRangeNonbondeds->updateAfterPartition(*md);
1012 // Allocate or re-size GPU halo exchange object, if necessary
1013 if (bNS && simulationWork.havePpDomainDecomposition && simulationWork.useGpuHaloExchange)
1015 GMX_RELEASE_ASSERT(fr->deviceStreamManager != nullptr,
1016 "GPU device manager has to be initialized to use GPU "
1017 "version of halo exchange.");
1018 constructGpuHaloExchange(mdlog, *cr, *fr->deviceStreamManager, wcycle);
1021 if (MASTER(cr) && do_log)
1023 gmx::EnergyOutput::printHeader(
1024 fplog, step, t); /* can we improve the information printed here? */
1027 if (ir->efep != FreeEnergyPerturbationType::No)
1029 update_mdatoms(mdAtoms->mdatoms(), state->lambda[FreeEnergyPerturbationCouplingType::Mass]);
1034 /* We need the kinetic energy at minus the half step for determining
1035 * the full step kinetic energy and possibly for T-coupling.*/
1036 /* This may not be quite working correctly yet . . . . */
1037 int cglo_flags = CGLO_GSTAT | CGLO_TEMPERATURE;
1038 compute_globals(gstat,
1043 makeConstArrayRef(state->x),
1044 makeConstArrayRef(state->v),
1060 &observablesReducer);
1062 clear_mat(force_vir);
1064 checkpointHandler->decideIfCheckpointingThisStep(bNS, bFirstStep, bLastStep);
1066 /* Determine the energy and pressure:
1067 * at nstcalcenergy steps and at energy output steps (set below).
1069 if (EI_VV(ir->eI) && (!bInitStep))
1071 bCalcEnerStep = do_per_step(step, ir->nstcalcenergy);
1072 bCalcVir = bCalcEnerStep
1073 || (ir->epc != PressureCoupling::No
1074 && (do_per_step(step, ir->nstpcouple) || do_per_step(step - 1, ir->nstpcouple)));
1078 bCalcEnerStep = do_per_step(step, ir->nstcalcenergy);
1079 bCalcVir = bCalcEnerStep
1080 || (ir->epc != PressureCoupling::No && do_per_step(step, ir->nstpcouple));
1082 bCalcEner = bCalcEnerStep;
1084 do_ene = (do_per_step(step, ir->nstenergy) || bLastStep);
1086 if (do_ene || do_log || bDoReplEx)
1092 // bCalcEner is only here for when the last step is not a mulitple of nstfep
1093 const bool computeDHDL = ((ir->efep != FreeEnergyPerturbationType::No || ir->bSimTemp)
1094 && (do_per_step(step, nstfep) || bCalcEner));
1096 /* Do we need global communication ? */
1097 bGStat = (bCalcVir || bCalcEner || bStopCM || do_per_step(step, nstglobalcomm)
1098 || (EI_VV(ir->eI) && inputrecNvtTrotter(ir) && do_per_step(step - 1, nstglobalcomm)));
1100 force_flags = (GMX_FORCE_STATECHANGED | ((inputrecDynamicBox(ir)) ? GMX_FORCE_DYNAMICBOX : 0)
1101 | GMX_FORCE_ALLFORCES | (bCalcVir ? GMX_FORCE_VIRIAL : 0)
1102 | (bCalcEner ? GMX_FORCE_ENERGY : 0) | (computeDHDL ? GMX_FORCE_DHDL : 0));
1103 if (simulationWork.useMts && !do_per_step(step, ir->nstfout))
1105 // TODO: merge this with stepWork.useOnlyMtsCombinedForceBuffer
1106 force_flags |= GMX_FORCE_DO_NOT_NEED_NORMAL_FORCE;
1111 /* Now is the time to relax the shells */
1112 relax_shell_flexcon(fplog,
1115 mdrunOptions.verbose,
1127 state->x.arrayRefWithPadding(),
1128 state->v.arrayRefWithPadding(),
1135 fr->longRangeNonbondeds.get(),
1144 ddBalanceRegionHandler);
1148 /* The AWH history need to be saved _before_ doing force calculations where the AWH bias
1149 is updated (or the AWH update will be performed twice for one step when continuing).
1150 It would be best to call this update function from do_md_trajectory_writing but that
1151 would occur after do_force. One would have to divide the update_awh function into one
1152 function applying the AWH force and one doing the AWH bias update. The update AWH
1153 bias function could then be called after do_md_trajectory_writing (then containing
1154 update_awh_history). The checkpointing will in the future probably moved to the start
1155 of the md loop which will rid of this issue. */
1156 if (awh && checkpointHandler->isCheckpointingStep() && MASTER(cr))
1158 awh->updateHistory(state_global->awhHistory.get());
1161 /* The coordinates (x) are shifted (to get whole molecules)
1163 * This is parallellized as well, and does communication too.
1164 * Check comments in sim_util.c
1179 state->x.arrayRefWithPadding(),
1191 ed ? ed->getLegacyED() : nullptr,
1192 fr->longRangeNonbondeds.get(),
1193 (bNS ? GMX_FORCE_NS : 0) | force_flags,
1194 ddBalanceRegionHandler);
1197 // VV integrators do not need the following velocity half step
1198 // if it is the first step after starting from a checkpoint.
1199 // That is, the half step is needed on all other steps, and
1200 // also the first step when starting from a .tpr file.
1203 integrateVVFirstStep(step,
1217 &observablesReducer,
1235 &saved_conserved_quantity,
1244 if (vsite != nullptr && needVirtualVelocitiesThisStep)
1246 // Positions were calculated earlier
1247 wallcycle_start(wcycle, WallCycleCounter::VsiteConstr);
1248 vsite->construct(state->x, state->v, state->box, VSiteOperation::Velocities);
1249 wallcycle_stop(wcycle, WallCycleCounter::VsiteConstr);
1253 /* ######## END FIRST UPDATE STEP ############## */
1254 /* ######## If doing VV, we now have v(dt) ###### */
1257 /* perform extended ensemble sampling in lambda - we don't
1258 actually move to the new state before outputting
1259 statistics, but if performing simulated tempering, we
1260 do update the velocities and the tau_t. */
1261 // TODO: Avoid changing inputrec (#3854)
1262 // Simulated tempering updates the reference temperature.
1263 // Expanded ensemble without simulated tempering does not change the inputrec.
1264 auto* nonConstInputrec = const_cast<t_inputrec*>(inputrec);
1265 lamnew = ExpandedEnsembleDynamics(fplog,
1273 state->v.rvec_array(),
1275 md->cTC ? gmx::arrayRefFromArray(md->cTC, md->nr)
1276 : gmx::ArrayRef<const unsigned short>());
1277 /* history is maintained in state->dfhist, but state_global is what is sent to trajectory and log output */
1280 copy_df_history(state_global->dfhist, state->dfhist);
1284 // Copy coordinate from the GPU for the output/checkpointing if the update is offloaded and
1285 // coordinates have not already been copied for i) search or ii) CPU force tasks.
1286 if (useGpuForUpdate && !bNS && !runScheduleWork->domainWork.haveCpuLocalForceWork
1287 && (do_per_step(step, ir->nstxout) || do_per_step(step, ir->nstxout_compressed)
1288 || checkpointHandler->isCheckpointingStep()))
1290 stateGpu->copyCoordinatesFromGpu(state->x, AtomLocality::Local);
1291 stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
1293 // Copy velocities if needed for the output/checkpointing.
1294 // NOTE: Copy on the search steps is done at the beginning of the step.
1295 if (useGpuForUpdate && !bNS
1296 && (do_per_step(step, ir->nstvout) || checkpointHandler->isCheckpointingStep()))
1298 stateGpu->copyVelocitiesFromGpu(state->v, AtomLocality::Local);
1299 stateGpu->waitVelocitiesReadyOnHost(AtomLocality::Local);
1301 // Copy forces for the output if the forces were reduced on the GPU (not the case on virial steps)
1302 // and update is offloaded hence forces are kept on the GPU for update and have not been
1303 // already transferred in do_force().
1304 // TODO: There should be an improved, explicit mechanism that ensures this copy is only executed
1305 // when the forces are ready on the GPU -- the same synchronizer should be used as the one
1306 // prior to GPU update.
1307 // TODO: When the output flags will be included in step workload, this copy can be combined with the
1308 // copy call in do_force(...).
1309 // NOTE: The forces should not be copied here if the vsites are present, since they were modified
1310 // on host after the D2H copy in do_force(...).
1311 if (runScheduleWork->stepWork.useGpuFBufferOps && (simulationWork.useGpuUpdate && !vsite)
1312 && do_per_step(step, ir->nstfout))
1314 stateGpu->copyForcesFromGpu(f.view().force(), AtomLocality::Local);
1315 stateGpu->waitForcesReadyOnHost(AtomLocality::Local);
1317 /* Now we have the energies and forces corresponding to the
1318 * coordinates at time t. We must output all of this before
1321 do_md_trajectory_writing(fplog,
1338 checkpointHandler->isCheckpointingStep(),
1341 mdrunOptions.writeConfout,
1343 /* Check if IMD step and do IMD communication, if bIMD is TRUE. */
1344 bInteractiveMDstep = imdSession->run(step, bNS, state->box, state->x, t);
1346 /* kludge -- virial is lost with restart for MTTK NPT control. Must reload (saved earlier). */
1347 if (startingBehavior != StartingBehavior::NewSimulation && bFirstStep
1348 && (inputrecNptTrotter(ir) || inputrecNphTrotter(ir)))
1350 copy_mat(state->svir_prev, shake_vir);
1351 copy_mat(state->fvir_prev, force_vir);
1354 stopHandler->setSignal();
1355 resetHandler->setSignal(walltime_accounting);
1357 if (bGStat || !PAR(cr))
1359 /* In parallel we only have to check for checkpointing in steps
1360 * where we do global communication,
1361 * otherwise the other nodes don't know.
1363 checkpointHandler->setSignal(walltime_accounting);
1366 /* ######### START SECOND UPDATE STEP ################# */
1368 /* at the start of step, randomize or scale the velocities ((if vv. Restriction of Andersen
1369 controlled in preprocessing */
1371 if (ETC_ANDERSEN(ir->etc)) /* keep this outside of update_tcouple because of the extra info required to pass */
1373 gmx_bool bIfRandomize;
1374 bIfRandomize = update_randomize_velocities(ir,
1378 md->cTC ? gmx::arrayRefFromArray(md->cTC, md->nr)
1379 : gmx::ArrayRef<const unsigned short>(),
1380 gmx::arrayRefFromArray(md->invmass, md->nr),
1384 /* if we have constraints, we have to remove the kinetic energy parallel to the bonds */
1385 if (constr && bIfRandomize)
1387 constrain_velocities(constr, do_log, do_ene, step, state, nullptr, false, nullptr);
1390 /* Box is changed in update() when we do pressure coupling,
1391 * but we should still use the old box for energy corrections and when
1392 * writing it to the energy file, so it matches the trajectory files for
1393 * the same timestep above. Make a copy in a separate array.
1395 copy_mat(state->box, lastbox);
1399 if (!useGpuForUpdate)
1401 wallcycle_start(wcycle, WallCycleCounter::Update);
1403 /* UPDATE PRESSURE VARIABLES IN TROTTER FORMULATION WITH CONSTRAINTS */
1413 md->cTC ? gmx::arrayRefFromArray(md->cTC, md->nr)
1414 : gmx::ArrayRef<const unsigned short>(),
1415 gmx::arrayRefFromArray(md->invmass, md->nr),
1418 TrotterSequence::Three);
1419 /* We can only do Berendsen coupling after we have summed
1420 * the kinetic energy or virial. Since the happens
1421 * in global_state after update, we should only do it at
1422 * step % nstlist = 1 with bGStatEveryStep=FALSE.
1427 update_tcouple(step,
1433 md->cTC ? gmx::arrayRefFromArray(md->cTC, md->nr)
1434 : gmx::ArrayRef<const unsigned short>());
1435 update_pcouple_before_coordinates(fplog, step, ir, state, pressureCouplingMu, M, bInitStep);
1438 /* With leap-frog type integrators we compute the kinetic energy
1439 * at a whole time step as the average of the half-time step kinetic
1440 * energies of two subsequent steps. Therefore we need to compute the
1441 * half step kinetic energy also if we need energies at the next step.
1443 const bool needHalfStepKineticEnergy =
1444 (!EI_VV(ir->eI) && (do_per_step(step + 1, nstglobalcomm) || step_rel + 1 == ir->nsteps));
1446 // Parrinello-Rahman requires the pressure to be availible before the update to compute
1447 // the velocity scaling matrix. Hence, it runs one step after the nstpcouple step.
1448 const bool doParrinelloRahman = (ir->epc == PressureCoupling::ParrinelloRahman
1449 && do_per_step(step + ir->nstpcouple - 1, ir->nstpcouple));
1453 GMX_ASSERT(!useGpuForUpdate, "GPU update is not supported with VVAK integrator.");
1455 integrateVVSecondStep(step,
1466 &observablesReducer,
1492 if (useGpuForUpdate)
1494 // On search steps, update handles to device vectors
1495 if (bNS && (bFirstStep || haveDDAtomOrdering(*cr) || bExchanged))
1497 integrator->set(stateGpu->getCoordinates(),
1498 stateGpu->getVelocities(),
1499 stateGpu->getForces(),
1503 // Copy data to the GPU after buffers might have being reinitialized
1504 /* The velocity copy is redundant if we had Center-of-Mass motion removed on
1505 * the previous step. We don't check that now. */
1506 stateGpu->copyVelocitiesToGpu(state->v, AtomLocality::Local);
1508 || (!runScheduleWork->stepWork.haveGpuPmeOnThisRank
1509 && !runScheduleWork->stepWork.useGpuXBufferOps))
1511 stateGpu->copyCoordinatesToGpu(state->x, AtomLocality::Local);
1515 if ((simulationWork.useGpuPme && simulationWork.useCpuPmePpCommunication)
1516 || (!runScheduleWork->stepWork.useGpuFBufferOps))
1518 // The PME forces were recieved to the host, and reduced on the CPU with the
1519 // rest of the forces computed on the GPU, so the final forces have to be copied
1520 // back to the GPU. Or the buffer ops were not offloaded this step, so the
1521 // forces are on the host and have to be copied
1522 stateGpu->copyForcesToGpu(f.view().force(), AtomLocality::Local);
1524 const bool doTemperatureScaling =
1525 (ir->etc != TemperatureCoupling::No
1526 && do_per_step(step + ir->nsttcouple - 1, ir->nsttcouple));
1528 // This applies Leap-Frog, LINCS and SETTLE in succession
1529 integrator->integrate(stateGpu->getLocalForcesReadyOnDeviceEvent(
1530 runScheduleWork->stepWork, runScheduleWork->simulationWork),
1535 doTemperatureScaling,
1538 ir->nstpcouple * ir->delta_t,
1543 /* With multiple time stepping we need to do an additional normal
1544 * update step to obtain the virial, as the actual MTS integration
1545 * using an acceleration where the slow forces are multiplied by mtsFactor.
1546 * Using that acceleration would result in a virial with the slow
1547 * force contribution would be a factor mtsFactor too large.
1549 if (simulationWork.useMts && bCalcVir && constr != nullptr)
1551 upd.update_for_constraint_virial(*ir,
1553 md->havePartiallyFrozenAtoms,
1554 gmx::arrayRefFromArray(md->invmass, md->nr),
1555 gmx::arrayRefFromArray(md->invMassPerDim, md->nr),
1557 f.view().forceWithPadding(),
1560 constrain_coordinates(constr,
1565 upd.xp()->arrayRefWithPadding(),
1571 ArrayRefWithPadding<const RVec> forceCombined =
1572 (simulationWork.useMts && step % ir->mtsLevels[1].stepFactor == 0)
1573 ? f.view().forceMtsCombinedWithPadding()
1574 : f.view().forceWithPadding();
1575 upd.update_coords(*ir,
1578 md->havePartiallyFrozenAtoms,
1579 gmx::arrayRefFromArray(md->ptype, md->nr),
1580 gmx::arrayRefFromArray(md->invmass, md->nr),
1581 gmx::arrayRefFromArray(md->invMassPerDim, md->nr),
1591 wallcycle_stop(wcycle, WallCycleCounter::Update);
1593 constrain_coordinates(constr,
1598 upd.xp()->arrayRefWithPadding(),
1600 bCalcVir && !simulationWork.useMts,
1603 upd.update_sd_second_half(*ir,
1607 gmx::arrayRefFromArray(md->ptype, md->nr),
1608 gmx::arrayRefFromArray(md->invmass, md->nr),
1617 *ir, md->havePartiallyFrozenAtoms, md->homenr, state, wcycle, constr != nullptr);
1620 if (ir->bPull && ir->pull->bSetPbcRefToPrevStepCOM)
1622 updatePrevStepPullCom(pull_work, state->pull_com_prev_step);
1625 enerd->term[F_DVDL_CONSTR] += dvdl_constr;
1628 /* ############## IF NOT VV, Calculate globals HERE ############ */
1629 /* With Leap-Frog we can skip compute_globals at
1630 * non-communication steps, but we need to calculate
1631 * the kinetic energy one step before communication.
1634 // Organize to do inter-simulation signalling on steps if
1635 // and when algorithms require it.
1636 const bool doInterSimSignal = (simulationsShareState && do_per_step(step, nstSignalComm));
1638 if (useGpuForUpdate)
1640 const bool coordinatesRequiredForStopCM =
1641 bStopCM && (bGStat || needHalfStepKineticEnergy || doInterSimSignal)
1644 // Copy coordinates when needed to stop the CM motion or for replica exchange
1645 if (coordinatesRequiredForStopCM || bDoReplEx)
1647 stateGpu->copyCoordinatesFromGpu(state->x, AtomLocality::Local);
1648 stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
1651 // Copy velocities back to the host if:
1652 // - Globals are computed this step (includes the energy output steps).
1653 // - Temperature is needed for the next step.
1654 // - This is a replica exchange step (even though we will only need
1655 // the velocities if an exchange succeeds)
1656 if (bGStat || needHalfStepKineticEnergy || bDoReplEx)
1658 stateGpu->copyVelocitiesFromGpu(state->v, AtomLocality::Local);
1659 stateGpu->waitVelocitiesReadyOnHost(AtomLocality::Local);
1663 if (bGStat || needHalfStepKineticEnergy || doInterSimSignal)
1665 // Since we're already communicating at this step, we
1666 // can propagate intra-simulation signals. Note that
1667 // check_nstglobalcomm has the responsibility for
1668 // choosing the value of nstglobalcomm that is one way
1669 // bGStat becomes true, so we can't get into a
1670 // situation where e.g. checkpointing can't be
1672 bool doIntraSimSignal = true;
1673 SimulationSignaller signaller(&signals, cr, ms, doInterSimSignal, doIntraSimSignal);
1675 compute_globals(gstat,
1680 makeConstArrayRef(state->x),
1681 makeConstArrayRef(state->v),
1695 (bGStat ? CGLO_GSTAT : 0) | (!EI_VV(ir->eI) && bCalcEner ? CGLO_ENERGY : 0)
1696 | (!EI_VV(ir->eI) && bStopCM ? CGLO_STOPCM : 0)
1697 | (!EI_VV(ir->eI) ? CGLO_TEMPERATURE : 0)
1698 | (!EI_VV(ir->eI) ? CGLO_PRESSURE : 0) | CGLO_CONSTRAINT,
1700 &observablesReducer);
1701 if (!EI_VV(ir->eI) && bStopCM)
1703 process_and_stopcm_grp(
1704 fplog, &vcm, *md, makeArrayRef(state->x), makeArrayRef(state->v));
1705 inc_nrnb(nrnb, eNR_STOPCM, md->homenr);
1707 // TODO: The special case of removing CM motion should be dealt more gracefully
1708 if (useGpuForUpdate)
1710 stateGpu->copyCoordinatesToGpu(state->x, AtomLocality::Local);
1711 // Here we block until the H2D copy completes because event sync with the
1712 // force kernels that use the coordinates on the next steps is not implemented
1713 // (not because of a race on state->x being modified on the CPU while H2D is in progress).
1714 stateGpu->waitCoordinatesCopiedToDevice(AtomLocality::Local);
1715 // If the COM removal changed the velocities on the CPU, this has to be accounted for.
1716 if (vcm.mode != ComRemovalAlgorithm::No)
1718 stateGpu->copyVelocitiesToGpu(state->v, AtomLocality::Local);
1725 /* ############# END CALC EKIN AND PRESSURE ################# */
1727 /* Note: this is OK, but there are some numerical precision issues with using the convergence of
1728 the virial that should probably be addressed eventually. state->veta has better properies,
1729 but what we actually need entering the new cycle is the new shake_vir value. Ideally, we could
1730 generate the new shake_vir, but test the veta value for convergence. This will take some thought. */
1732 if (ir->efep != FreeEnergyPerturbationType::No && !EI_VV(ir->eI))
1734 /* Sum up the foreign energy and dK/dl terms for md and sd.
1735 Currently done every step so that dH/dl is correct in the .edr */
1736 accumulateKineticLambdaComponents(enerd, state->lambda, *ir->fepvals);
1739 bool scaleCoordinates = !useGpuForUpdate || bDoReplEx;
1740 update_pcouple_after_coordinates(fplog,
1744 md->cFREEZE ? gmx::arrayRefFromArray(md->cFREEZE, md->nr)
1745 : gmx::ArrayRef<const unsigned short>(),
1755 const bool doBerendsenPressureCoupling = (inputrec->epc == PressureCoupling::Berendsen
1756 && do_per_step(step, inputrec->nstpcouple));
1757 const bool doCRescalePressureCoupling = (inputrec->epc == PressureCoupling::CRescale
1758 && do_per_step(step, inputrec->nstpcouple));
1760 && (doBerendsenPressureCoupling || doCRescalePressureCoupling || doParrinelloRahman))
1762 integrator->scaleCoordinates(pressureCouplingMu);
1763 if (doCRescalePressureCoupling)
1765 matrix pressureCouplingInvMu;
1766 gmx::invertBoxMatrix(pressureCouplingMu, pressureCouplingInvMu);
1767 integrator->scaleVelocities(pressureCouplingInvMu);
1769 integrator->setPbc(PbcType::Xyz, state->box);
1772 /* ################# END UPDATE STEP 2 ################# */
1773 /* #### We now have r(t+dt) and v(t+dt/2) ############# */
1775 /* The coordinates (x) were unshifted in update */
1778 /* We will not sum ekinh_old,
1779 * so signal that we still have to do it.
1781 bSumEkinhOld = TRUE;
1786 /* ######### BEGIN PREPARING EDR OUTPUT ########### */
1788 /* use the directly determined last velocity, not actually the averaged half steps */
1789 if (bTrotter && ir->eI == IntegrationAlgorithm::VV)
1791 enerd->term[F_EKIN] = last_ekin;
1793 enerd->term[F_ETOT] = enerd->term[F_EPOT] + enerd->term[F_EKIN];
1795 if (integratorHasConservedEnergyQuantity(ir))
1799 enerd->term[F_ECONSERVED] = enerd->term[F_ETOT] + saved_conserved_quantity;
1803 enerd->term[F_ECONSERVED] = enerd->term[F_ETOT] + NPT_energy(ir, state, &MassQ);
1806 /* ######### END PREPARING EDR OUTPUT ########### */
1812 if (fplog && do_log && bDoExpanded)
1814 /* only needed if doing expanded ensemble */
1815 PrintFreeEnergyInfoToFile(fplog,
1817 ir->expandedvals.get(),
1818 ir->bSimTemp ? ir->simtempvals.get() : nullptr,
1819 state_global->dfhist,
1826 const bool outputDHDL = (computeDHDL && do_per_step(step, ir->fepvals->nstdhdl));
1828 energyOutput.addDataAtEnergyStep(outputDHDL,
1834 ir->expandedvals.get(),
1836 PTCouplingArrays{ state->boxv,
1837 state->nosehoover_xi,
1838 state->nosehoover_vxi,
1840 state->nhpres_vxi },
1850 energyOutput.recordNonEnergyStep();
1853 gmx_bool do_dr = do_per_step(step, ir->nstdisreout);
1854 gmx_bool do_or = do_per_step(step, ir->nstorireout);
1856 if (doSimulatedAnnealing)
1858 gmx::EnergyOutput::printAnnealingTemperatures(
1859 do_log ? fplog : nullptr, groups, &(ir->opts));
1861 if (do_log || do_ene || do_dr || do_or)
1863 energyOutput.printStepToEnergyFile(mdoutf_get_fp_ene(outf),
1867 do_log ? fplog : nullptr,
1873 if (do_log && ir->bDoAwh && awh->hasFepLambdaDimension())
1875 const bool isInitialOutput = false;
1876 printLambdaStateToLog(fplog, state->lambda, isInitialOutput);
1881 pull_print_output(pull_work, step, t);
1884 if (do_per_step(step, ir->nstlog))
1886 if (fflush(fplog) != 0)
1888 gmx_fatal(FARGS, "Cannot flush logfile - maybe you are out of disk space?");
1894 /* Have to do this part _after_ outputting the logfile and the edr file */
1895 /* Gets written into the state at the beginning of next loop*/
1896 state->fep_state = lamnew;
1898 else if (ir->bDoAwh && awh->needForeignEnergyDifferences(step))
1900 state->fep_state = awh->fepLambdaState();
1902 /* Print the remaining wall clock time for the run */
1903 if (isMasterSimMasterRank(ms, MASTER(cr)) && (do_verbose || gmx_got_usr_signal()) && !bPMETunePrinting)
1907 fprintf(stderr, "\n");
1909 print_time(stderr, walltime_accounting, step, ir, cr);
1912 /* Ion/water position swapping.
1913 * Not done in last step since trajectory writing happens before this call
1914 * in the MD loop and exchanges would be lost anyway. */
1915 bNeedRepartition = FALSE;
1916 if ((ir->eSwapCoords != SwapType::No) && (step > 0) && !bLastStep
1917 && do_per_step(step, ir->swap->nstswap))
1919 bNeedRepartition = do_swapcoords(cr,
1925 as_rvec_array(state->x.data()),
1927 MASTER(cr) && mdrunOptions.verbose,
1930 if (bNeedRepartition && haveDDAtomOrdering(*cr))
1932 dd_collect_state(cr->dd, state, state_global);
1936 /* Replica exchange */
1940 bExchanged = replica_exchange(fplog, cr, ms, repl_ex, state_global, enerd, state, step, t);
1943 if ((bExchanged || bNeedRepartition) && haveDDAtomOrdering(*cr))
1945 dd_partition_system(fplog,
1966 upd.updateAfterPartition(state->natoms,
1967 md->cFREEZE ? gmx::arrayRefFromArray(md->cFREEZE, md->nr)
1968 : gmx::ArrayRef<const unsigned short>(),
1969 md->cTC ? gmx::arrayRefFromArray(md->cTC, md->nr)
1970 : gmx::ArrayRef<const unsigned short>());
1971 fr->longRangeNonbondeds->updateAfterPartition(*md);
1977 /* ####### SET VARIABLES FOR NEXT ITERATION IF THEY STILL NEED IT ###### */
1978 /* With all integrators, except VV, we need to retain the pressure
1979 * at the current step for coupling at the next step.
1981 if ((state->flags & enumValueToBitMask(StateEntry::PressurePrevious))
1982 && (bGStatEveryStep || (ir->nstpcouple > 0 && step % ir->nstpcouple == 0)))
1984 /* Store the pressure in t_state for pressure coupling
1985 * at the next MD step.
1987 copy_mat(pres, state->pres_prev);
1990 /* ####### END SET VARIABLES FOR NEXT ITERATION ###### */
1992 if ((membed != nullptr) && (!bLastStep))
1994 rescale_membed(step_rel, membed, as_rvec_array(state_global->x.data()));
1997 cycles = wallcycle_stop(wcycle, WallCycleCounter::Step);
1998 if (haveDDAtomOrdering(*cr) && wcycle)
2000 dd_cycles_add(cr->dd, cycles, ddCyclStep);
2003 /* increase the MD step number */
2006 observablesReducer.markAsReadyToReduce();
2011 fcReportProgress(ir->nsteps + ir->init_step, step);
2015 resetHandler->resetCounters(
2016 step, step_rel, mdlog, fplog, cr, fr->nbv.get(), nrnb, fr->pmedata, pme_loadbal, wcycle, walltime_accounting);
2018 /* If bIMD is TRUE, the master updates the IMD energy record and sends positions to VMD client */
2019 imdSession->updateEnergyRecordAndSendPositionsAndEnergies(bInteractiveMDstep, step, bCalcEner);
2021 /* End of main MD loop */
2023 /* Closing TNG files can include compressing data. Therefore it is good to do that
2024 * before stopping the time measurements. */
2025 mdoutf_tng_close(outf);
2027 /* Stop measuring walltime */
2028 walltime_accounting_end_time(walltime_accounting);
2030 if (simulationWork.haveSeparatePmeRank)
2032 /* Tell the PME only node to finish */
2033 gmx_pme_send_finish(cr);
2038 if (ir->nstcalcenergy > 0)
2040 energyOutput.printEnergyConservation(fplog, ir->simulation_part, EI_MD(ir->eI));
2042 gmx::EnergyOutput::printAnnealingTemperatures(fplog, groups, &(ir->opts));
2043 energyOutput.printAverages(fplog, groups);
2050 pme_loadbal_done(pme_loadbal, fplog, mdlog, fr->nbv->useGpu());
2053 done_shellfc(fplog, shellfc, step_rel);
2055 if (useReplicaExchange && MASTER(cr))
2057 print_replica_exchange_statistics(fplog, repl_ex);
2060 walltime_accounting_set_nsteps_done(walltime_accounting, step_rel);
2062 global_stat_destroy(gstat);