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 MD runner routine calling all integrators.
41 * \author David van der Spoel <david.vanderspoel@icm.uu.se>
42 * \ingroup module_mdrun
59 #include "gromacs/commandline/filenm.h"
60 #include "gromacs/domdec/builder.h"
61 #include "gromacs/domdec/domdec.h"
62 #include "gromacs/domdec/domdec_struct.h"
63 #include "gromacs/domdec/gpuhaloexchange.h"
64 #include "gromacs/domdec/localatomsetmanager.h"
65 #include "gromacs/domdec/partition.h"
66 #include "gromacs/ewald/ewald_utils.h"
67 #include "gromacs/ewald/pme.h"
68 #include "gromacs/ewald/pme_gpu_program.h"
69 #include "gromacs/ewald/pme_only.h"
70 #include "gromacs/ewald/pme_pp_comm_gpu.h"
71 #include "gromacs/fileio/checkpoint.h"
72 #include "gromacs/fileio/gmxfio.h"
73 #include "gromacs/fileio/oenv.h"
74 #include "gromacs/fileio/tpxio.h"
75 #include "gromacs/gmxlib/network.h"
76 #include "gromacs/gmxlib/nrnb.h"
77 #include "gromacs/gpu_utils/device_stream_manager.h"
78 #include "gromacs/hardware/cpuinfo.h"
79 #include "gromacs/hardware/detecthardware.h"
80 #include "gromacs/hardware/device_management.h"
81 #include "gromacs/hardware/hardwaretopology.h"
82 #include "gromacs/hardware/printhardware.h"
83 #include "gromacs/imd/imd.h"
84 #include "gromacs/listed_forces/disre.h"
85 #include "gromacs/listed_forces/gpubonded.h"
86 #include "gromacs/listed_forces/listed_forces.h"
87 #include "gromacs/listed_forces/orires.h"
88 #include "gromacs/math/functions.h"
89 #include "gromacs/math/utilities.h"
90 #include "gromacs/math/vec.h"
91 #include "gromacs/mdlib/boxdeformation.h"
92 #include "gromacs/mdlib/broadcaststructs.h"
93 #include "gromacs/mdlib/calc_verletbuf.h"
94 #include "gromacs/mdlib/dispersioncorrection.h"
95 #include "gromacs/mdlib/enerdata_utils.h"
96 #include "gromacs/mdlib/force.h"
97 #include "gromacs/mdlib/forcerec.h"
98 #include "gromacs/mdlib/gmx_omp_nthreads.h"
99 #include "gromacs/mdlib/gpuforcereduction.h"
100 #include "gromacs/mdlib/makeconstraints.h"
101 #include "gromacs/mdlib/md_support.h"
102 #include "gromacs/mdlib/mdatoms.h"
103 #include "gromacs/mdlib/sighandler.h"
104 #include "gromacs/mdlib/stophandler.h"
105 #include "gromacs/mdlib/tgroup.h"
106 #include "gromacs/mdlib/updategroups.h"
107 #include "gromacs/mdlib/vsite.h"
108 #include "gromacs/mdrun/mdmodules.h"
109 #include "gromacs/mdrun/simulationcontext.h"
110 #include "gromacs/mdrun/simulationinput.h"
111 #include "gromacs/mdrun/simulationinputhandle.h"
112 #include "gromacs/mdrunutility/handlerestart.h"
113 #include "gromacs/mdrunutility/logging.h"
114 #include "gromacs/mdrunutility/multisim.h"
115 #include "gromacs/mdrunutility/printtime.h"
116 #include "gromacs/mdrunutility/threadaffinity.h"
117 #include "gromacs/mdtypes/checkpointdata.h"
118 #include "gromacs/mdtypes/commrec.h"
119 #include "gromacs/mdtypes/enerdata.h"
120 #include "gromacs/mdtypes/fcdata.h"
121 #include "gromacs/mdtypes/forcerec.h"
122 #include "gromacs/mdtypes/group.h"
123 #include "gromacs/mdtypes/inputrec.h"
124 #include "gromacs/mdtypes/interaction_const.h"
125 #include "gromacs/mdtypes/md_enums.h"
126 #include "gromacs/mdtypes/mdatom.h"
127 #include "gromacs/mdtypes/mdrunoptions.h"
128 #include "gromacs/mdtypes/observableshistory.h"
129 #include "gromacs/mdtypes/simulation_workload.h"
130 #include "gromacs/mdtypes/state.h"
131 #include "gromacs/mdtypes/state_propagator_data_gpu.h"
132 #include "gromacs/modularsimulator/modularsimulator.h"
133 #include "gromacs/nbnxm/gpu_data_mgmt.h"
134 #include "gromacs/nbnxm/nbnxm.h"
135 #include "gromacs/nbnxm/pairlist_tuning.h"
136 #include "gromacs/pbcutil/pbc.h"
137 #include "gromacs/pulling/output.h"
138 #include "gromacs/pulling/pull.h"
139 #include "gromacs/pulling/pull_rotation.h"
140 #include "gromacs/restraint/manager.h"
141 #include "gromacs/restraint/restraintmdmodule.h"
142 #include "gromacs/restraint/restraintpotential.h"
143 #include "gromacs/swap/swapcoords.h"
144 #include "gromacs/taskassignment/decidegpuusage.h"
145 #include "gromacs/taskassignment/decidesimulationworkload.h"
146 #include "gromacs/taskassignment/resourcedivision.h"
147 #include "gromacs/taskassignment/taskassignment.h"
148 #include "gromacs/taskassignment/usergpuids.h"
149 #include "gromacs/timing/gpu_timing.h"
150 #include "gromacs/timing/wallcycle.h"
151 #include "gromacs/timing/wallcyclereporting.h"
152 #include "gromacs/topology/mtop_util.h"
153 #include "gromacs/trajectory/trajectoryframe.h"
154 #include "gromacs/utility/basenetwork.h"
155 #include "gromacs/utility/cstringutil.h"
156 #include "gromacs/utility/exceptions.h"
157 #include "gromacs/utility/fatalerror.h"
158 #include "gromacs/utility/filestream.h"
159 #include "gromacs/utility/gmxassert.h"
160 #include "gromacs/utility/gmxmpi.h"
161 #include "gromacs/utility/keyvaluetree.h"
162 #include "gromacs/utility/logger.h"
163 #include "gromacs/utility/loggerbuilder.h"
164 #include "gromacs/utility/mdmodulenotification.h"
165 #include "gromacs/utility/physicalnodecommunicator.h"
166 #include "gromacs/utility/pleasecite.h"
167 #include "gromacs/utility/programcontext.h"
168 #include "gromacs/utility/smalloc.h"
169 #include "gromacs/utility/stringutil.h"
171 #include "isimulator.h"
172 #include "membedholder.h"
173 #include "replicaexchange.h"
174 #include "simulatorbuilder.h"
180 /*! \brief Manage any development feature flag variables encountered
182 * The use of dev features indicated by environment variables is
183 * logged in order to ensure that runs with such features enabled can
184 * be identified from their log and standard output. Any cross
185 * dependencies are also checked, and if unsatisfied, a fatal error
188 * Note that some development features overrides are applied already here:
189 * the GPU communication flags are set to false in non-tMPI and non-CUDA builds.
191 * \param[in] mdlog Logger object.
192 * \param[in] useGpuForNonbonded True if the nonbonded task is offloaded in this run.
193 * \param[in] pmeRunMode The PME run mode for this run
194 * \returns The object populated with development feature flags.
196 static DevelopmentFeatureFlags manageDevelopmentFeatures(const gmx::MDLogger& mdlog,
197 const bool useGpuForNonbonded,
198 const PmeRunMode pmeRunMode)
200 DevelopmentFeatureFlags devFlags;
202 // Some builds of GCC 5 give false positive warnings that these
203 // getenv results are ignored when clearly they are used.
204 #pragma GCC diagnostic push
205 #pragma GCC diagnostic ignored "-Wunused-result"
207 devFlags.enableGpuBufferOps =
208 GMX_GPU_CUDA && useGpuForNonbonded && (getenv("GMX_USE_GPU_BUFFER_OPS") != nullptr);
209 devFlags.enableGpuHaloExchange = GMX_GPU_CUDA && GMX_THREAD_MPI && getenv("GMX_GPU_DD_COMMS") != nullptr;
210 devFlags.forceGpuUpdateDefault = (getenv("GMX_FORCE_UPDATE_DEFAULT_GPU") != nullptr) || GMX_FAHCORE;
211 devFlags.enableGpuPmePPComm =
212 GMX_GPU_CUDA && GMX_THREAD_MPI && getenv("GMX_GPU_PME_PP_COMMS") != nullptr;
214 #pragma GCC diagnostic pop
216 if (devFlags.enableGpuBufferOps)
218 GMX_LOG(mdlog.warning)
220 .appendTextFormatted(
221 "This run uses the 'GPU buffer ops' feature, enabled by the "
222 "GMX_USE_GPU_BUFFER_OPS environment variable.");
225 if (devFlags.forceGpuUpdateDefault)
227 GMX_LOG(mdlog.warning)
229 .appendTextFormatted(
230 "This run will default to '-update gpu' as requested by the "
231 "GMX_FORCE_UPDATE_DEFAULT_GPU environment variable. GPU update with domain "
232 "decomposition lacks substantial testing and should be used with caution.");
235 if (devFlags.enableGpuHaloExchange)
237 if (useGpuForNonbonded)
239 if (!devFlags.enableGpuBufferOps)
241 GMX_LOG(mdlog.warning)
243 .appendTextFormatted(
244 "Enabling GPU buffer operations required by GMX_GPU_DD_COMMS "
245 "(equivalent with GMX_USE_GPU_BUFFER_OPS=1).");
246 devFlags.enableGpuBufferOps = true;
248 GMX_LOG(mdlog.warning)
250 .appendTextFormatted(
251 "This run has requested the 'GPU halo exchange' feature, enabled by "
253 "GMX_GPU_DD_COMMS environment variable.");
257 GMX_LOG(mdlog.warning)
259 .appendTextFormatted(
260 "GMX_GPU_DD_COMMS environment variable detected, but the 'GPU "
261 "halo exchange' feature will not be enabled as nonbonded interactions "
262 "are not offloaded.");
263 devFlags.enableGpuHaloExchange = false;
267 if (devFlags.enableGpuPmePPComm)
269 if (pmeRunMode == PmeRunMode::GPU)
271 if (!devFlags.enableGpuBufferOps)
273 GMX_LOG(mdlog.warning)
275 .appendTextFormatted(
276 "Enabling GPU buffer operations required by GMX_GPU_PME_PP_COMMS "
277 "(equivalent with GMX_USE_GPU_BUFFER_OPS=1).");
278 devFlags.enableGpuBufferOps = true;
280 GMX_LOG(mdlog.warning)
282 .appendTextFormatted(
283 "This run uses the 'GPU PME-PP communications' feature, enabled "
284 "by the GMX_GPU_PME_PP_COMMS environment variable.");
288 std::string clarification;
289 if (pmeRunMode == PmeRunMode::Mixed)
292 "PME FFT and gather are not offloaded to the GPU (PME is running in mixed "
297 clarification = "PME is not offloaded to the GPU.";
299 GMX_LOG(mdlog.warning)
302 "GMX_GPU_PME_PP_COMMS environment variable detected, but the "
303 "'GPU PME-PP communications' feature was not enabled as "
305 devFlags.enableGpuPmePPComm = false;
312 /*! \brief Barrier for safe simultaneous thread access to mdrunner data
314 * Used to ensure that the master thread does not modify mdrunner during copy
315 * on the spawned threads. */
316 static void threadMpiMdrunnerAccessBarrier()
319 MPI_Barrier(MPI_COMM_WORLD);
323 Mdrunner Mdrunner::cloneOnSpawnedThread() const
325 auto newRunner = Mdrunner(std::make_unique<MDModules>());
327 // All runners in the same process share a restraint manager resource because it is
328 // part of the interface to the client code, which is associated only with the
329 // original thread. Handles to the same resources can be obtained by copy.
331 newRunner.restraintManager_ = std::make_unique<RestraintManager>(*restraintManager_);
334 // Copy members of master runner.
335 // \todo Replace with builder when Simulation context and/or runner phases are better defined.
336 // Ref https://gitlab.com/gromacs/gromacs/-/issues/2587 and https://gitlab.com/gromacs/gromacs/-/issues/2375
337 newRunner.hw_opt = hw_opt;
338 newRunner.filenames = filenames;
340 newRunner.hwinfo_ = hwinfo_;
341 newRunner.oenv = oenv;
342 newRunner.mdrunOptions = mdrunOptions;
343 newRunner.domdecOptions = domdecOptions;
344 newRunner.nbpu_opt = nbpu_opt;
345 newRunner.pme_opt = pme_opt;
346 newRunner.pme_fft_opt = pme_fft_opt;
347 newRunner.bonded_opt = bonded_opt;
348 newRunner.update_opt = update_opt;
349 newRunner.nstlist_cmdline = nstlist_cmdline;
350 newRunner.replExParams = replExParams;
351 newRunner.pforce = pforce;
352 // Give the spawned thread the newly created valid communicator
353 // for the simulation.
354 newRunner.libraryWorldCommunicator = MPI_COMM_WORLD;
355 newRunner.simulationCommunicator = MPI_COMM_WORLD;
357 newRunner.startingBehavior = startingBehavior;
358 newRunner.stopHandlerBuilder_ = std::make_unique<StopHandlerBuilder>(*stopHandlerBuilder_);
359 newRunner.inputHolder_ = inputHolder_;
361 threadMpiMdrunnerAccessBarrier();
366 /*! \brief The callback used for running on spawned threads.
368 * Obtains the pointer to the master mdrunner object from the one
369 * argument permitted to the thread-launch API call, copies it to make
370 * a new runner for this thread, reinitializes necessary data, and
371 * proceeds to the simulation. */
372 static void mdrunner_start_fn(const void* arg)
376 auto masterMdrunner = reinterpret_cast<const gmx::Mdrunner*>(arg);
377 /* copy the arg list to make sure that it's thread-local. This
378 doesn't copy pointed-to items, of course; fnm, cr and fplog
379 are reset in the call below, all others should be const. */
380 gmx::Mdrunner mdrunner = masterMdrunner->cloneOnSpawnedThread();
383 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
387 void Mdrunner::spawnThreads(int numThreadsToLaunch)
390 /* now spawn new threads that start mdrunner_start_fn(), while
391 the main thread returns. Thread affinity is handled later. */
392 if (tMPI_Init_fn(TRUE, numThreadsToLaunch, TMPI_AFFINITY_NONE, mdrunner_start_fn, static_cast<const void*>(this))
395 GMX_THROW(gmx::InternalError("Failed to spawn thread-MPI threads"));
398 // Give the master thread the newly created valid communicator for
400 libraryWorldCommunicator = MPI_COMM_WORLD;
401 simulationCommunicator = MPI_COMM_WORLD;
402 threadMpiMdrunnerAccessBarrier();
404 GMX_UNUSED_VALUE(numThreadsToLaunch);
405 GMX_UNUSED_VALUE(mdrunner_start_fn);
411 /*! \brief Initialize variables for Verlet scheme simulation */
412 static void prepare_verlet_scheme(FILE* fplog,
416 const gmx_mtop_t* mtop,
418 bool makeGpuPairList,
419 const gmx::CpuInfo& cpuinfo)
421 // We checked the cut-offs in grompp, but double-check here.
422 // We have PME+LJcutoff kernels for rcoulomb>rvdw.
423 if (EEL_PME_EWALD(ir->coulombtype) && ir->vdwtype == VanDerWaalsType::Cut)
425 GMX_RELEASE_ASSERT(ir->rcoulomb >= ir->rvdw,
426 "With Verlet lists and PME we should have rcoulomb>=rvdw");
430 GMX_RELEASE_ASSERT(ir->rcoulomb == ir->rvdw,
431 "With Verlet lists and no PME rcoulomb and rvdw should be identical");
433 /* For NVE simulations, we will retain the initial list buffer */
434 if (EI_DYNAMICS(ir->eI) && ir->verletbuf_tol > 0
435 && !(EI_MD(ir->eI) && ir->etc == TemperatureCoupling::No))
437 /* Update the Verlet buffer size for the current run setup */
439 /* Here we assume SIMD-enabled kernels are being used. But as currently
440 * calc_verlet_buffer_size gives the same results for 4x8 and 4x4
441 * and 4x2 gives a larger buffer than 4x4, this is ok.
443 ListSetupType listType =
444 (makeGpuPairList ? ListSetupType::Gpu : ListSetupType::CpuSimdWhenSupported);
445 VerletbufListSetup listSetup = verletbufGetSafeListSetup(listType);
447 const real rlist_new =
448 calcVerletBufferSize(*mtop, det(box), *ir, ir->nstlist, ir->nstlist - 1, -1, listSetup);
450 if (rlist_new != ir->rlist)
452 if (fplog != nullptr)
455 "\nChanging rlist from %g to %g for non-bonded %dx%d atom kernels\n\n",
458 listSetup.cluster_size_i,
459 listSetup.cluster_size_j);
461 ir->rlist = rlist_new;
465 if (nstlist_cmdline > 0 && (!EI_DYNAMICS(ir->eI) || ir->verletbuf_tol <= 0))
468 "Can not set nstlist without %s",
469 !EI_DYNAMICS(ir->eI) ? "dynamics" : "verlet-buffer-tolerance");
472 if (EI_DYNAMICS(ir->eI))
474 /* Set or try nstlist values */
475 increaseNstlist(fplog, cr, ir, nstlist_cmdline, mtop, box, makeGpuPairList, cpuinfo);
479 /*! \brief Override the nslist value in inputrec
481 * with value passed on the command line (if any)
483 static void override_nsteps_cmdline(const gmx::MDLogger& mdlog, int64_t nsteps_cmdline, t_inputrec* ir)
487 /* override with anything else than the default -2 */
488 if (nsteps_cmdline > -2)
490 char sbuf_steps[STEPSTRSIZE];
491 char sbuf_msg[STRLEN];
493 ir->nsteps = nsteps_cmdline;
494 if (EI_DYNAMICS(ir->eI) && nsteps_cmdline != -1)
497 "Overriding nsteps with value passed on the command line: %s steps, %.3g ps",
498 gmx_step_str(nsteps_cmdline, sbuf_steps),
499 fabs(nsteps_cmdline * ir->delta_t));
504 "Overriding nsteps with value passed on the command line: %s steps",
505 gmx_step_str(nsteps_cmdline, sbuf_steps));
508 GMX_LOG(mdlog.warning).asParagraph().appendText(sbuf_msg);
510 else if (nsteps_cmdline < -2)
512 gmx_fatal(FARGS, "Invalid nsteps value passed on the command line: %" PRId64, nsteps_cmdline);
514 /* Do nothing if nsteps_cmdline == -2 */
520 /*! \brief Return whether GPU acceleration of nonbondeds is supported with the given settings.
522 * If not, and if a warning may be issued, logs a warning about
523 * falling back to CPU code. With thread-MPI, only the first
524 * call to this function should have \c issueWarning true. */
525 static bool gpuAccelerationOfNonbondedIsUseful(const MDLogger& mdlog, const t_inputrec& ir, bool issueWarning)
527 bool gpuIsUseful = true;
530 if (ir.opts.ngener - ir.nwall > 1)
532 /* The GPU code does not support more than one energy group.
533 * If the user requested GPUs explicitly, a fatal error is given later.
537 "Multiple energy groups is not implemented for GPUs, falling back to the CPU. "
538 "For better performance, run on the GPU without energy groups and then do "
539 "gmx mdrun -rerun option on the trajectory with an energy group .tpr file.";
545 warning = "TPI is not implemented for GPUs.";
548 if (!gpuIsUseful && issueWarning)
550 GMX_LOG(mdlog.warning).asParagraph().appendText(warning);
556 //! Initializes the logger for mdrun.
557 static gmx::LoggerOwner buildLogger(FILE* fplog, const bool isSimulationMasterRank)
559 gmx::LoggerBuilder builder;
560 if (fplog != nullptr)
562 builder.addTargetFile(gmx::MDLogger::LogLevel::Info, fplog);
564 if (isSimulationMasterRank)
566 builder.addTargetStream(gmx::MDLogger::LogLevel::Warning, &gmx::TextOutputFile::standardError());
568 return builder.build();
571 //! Make a TaskTarget from an mdrun argument string.
572 static TaskTarget findTaskTarget(const char* optionString)
574 TaskTarget returnValue = TaskTarget::Auto;
576 if (strncmp(optionString, "auto", 3) == 0)
578 returnValue = TaskTarget::Auto;
580 else if (strncmp(optionString, "cpu", 3) == 0)
582 returnValue = TaskTarget::Cpu;
584 else if (strncmp(optionString, "gpu", 3) == 0)
586 returnValue = TaskTarget::Gpu;
590 GMX_ASSERT(false, "Option string should have been checked for sanity already");
596 //! Finish run, aggregate data to print performance info.
597 static void finish_run(FILE* fplog,
598 const gmx::MDLogger& mdlog,
600 const t_inputrec& inputrec,
602 gmx_wallcycle_t wcycle,
603 gmx_walltime_accounting_t walltime_accounting,
604 nonbonded_verlet_t* nbv,
605 const gmx_pme_t* pme,
609 double nbfs = 0, mflop = 0;
610 double elapsed_time, elapsed_time_over_all_ranks, elapsed_time_over_all_threads,
611 elapsed_time_over_all_threads_over_all_ranks;
612 /* Control whether it is valid to print a report. Only the
613 simulation master may print, but it should not do so if the run
614 terminated e.g. before a scheduled reset step. This is
615 complicated by the fact that PME ranks are unaware of the
616 reason why they were sent a pmerecvqxFINISH. To avoid
617 communication deadlocks, we always do the communication for the
618 report, even if we've decided not to write the report, because
619 how long it takes to finish the run is not important when we've
620 decided not to report on the simulation performance.
622 Further, we only report performance for dynamical integrators,
623 because those are the only ones for which we plan to
624 consider doing any optimizations. */
625 bool printReport = EI_DYNAMICS(inputrec.eI) && SIMMASTER(cr);
627 if (printReport && !walltime_accounting_get_valid_finish(walltime_accounting))
629 GMX_LOG(mdlog.warning)
631 .appendText("Simulation ended prematurely, no performance report will be written.");
636 std::unique_ptr<t_nrnb> nrnbTotalStorage;
639 nrnbTotalStorage = std::make_unique<t_nrnb>();
640 nrnb_tot = nrnbTotalStorage.get();
642 MPI_Allreduce(nrnb->n, nrnb_tot->n, eNRNB, MPI_DOUBLE, MPI_SUM, cr->mpi_comm_mysim);
650 elapsed_time = walltime_accounting_get_time_since_reset(walltime_accounting);
651 elapsed_time_over_all_threads =
652 walltime_accounting_get_time_since_reset_over_all_threads(walltime_accounting);
656 /* reduce elapsed_time over all MPI ranks in the current simulation */
657 MPI_Allreduce(&elapsed_time, &elapsed_time_over_all_ranks, 1, MPI_DOUBLE, MPI_SUM, cr->mpi_comm_mysim);
658 elapsed_time_over_all_ranks /= cr->nnodes;
659 /* Reduce elapsed_time_over_all_threads over all MPI ranks in the
660 * current simulation. */
661 MPI_Allreduce(&elapsed_time_over_all_threads,
662 &elapsed_time_over_all_threads_over_all_ranks,
671 elapsed_time_over_all_ranks = elapsed_time;
672 elapsed_time_over_all_threads_over_all_ranks = elapsed_time_over_all_threads;
677 print_flop(fplog, nrnb_tot, &nbfs, &mflop);
680 if (thisRankHasDuty(cr, DUTY_PP) && DOMAINDECOMP(cr))
682 print_dd_statistics(cr, inputrec, fplog);
685 /* TODO Move the responsibility for any scaling by thread counts
686 * to the code that handled the thread region, so that there's a
687 * mechanism to keep cycle counting working during the transition
688 * to task parallelism. */
689 int nthreads_pp = gmx_omp_nthreads_get(emntNonbonded);
690 int nthreads_pme = gmx_omp_nthreads_get(emntPME);
691 wallcycle_scale_by_num_threads(
692 wcycle, thisRankHasDuty(cr, DUTY_PME) && !thisRankHasDuty(cr, DUTY_PP), nthreads_pp, nthreads_pme);
693 auto cycle_sum(wallcycle_sum(cr, wcycle));
697 auto nbnxn_gpu_timings =
698 (nbv != nullptr && nbv->useGpu()) ? Nbnxm::gpu_get_timings(nbv->gpu_nbv) : nullptr;
699 gmx_wallclock_gpu_pme_t pme_gpu_timings = {};
701 if (pme_gpu_task_enabled(pme))
703 pme_gpu_get_timings(pme, &pme_gpu_timings);
705 wallcycle_print(fplog,
711 elapsed_time_over_all_ranks,
717 if (EI_DYNAMICS(inputrec.eI))
719 delta_t = inputrec.delta_t;
725 elapsed_time_over_all_threads_over_all_ranks,
726 elapsed_time_over_all_ranks,
727 walltime_accounting_get_nsteps_done_since_reset(walltime_accounting),
735 elapsed_time_over_all_threads_over_all_ranks,
736 elapsed_time_over_all_ranks,
737 walltime_accounting_get_nsteps_done_since_reset(walltime_accounting),
745 int Mdrunner::mdrunner()
748 std::unique_ptr<t_forcerec> fr;
749 real ewaldcoeff_q = 0;
750 real ewaldcoeff_lj = 0;
751 int nChargePerturbed = -1, nTypePerturbed = 0;
752 gmx_wallcycle_t wcycle;
753 gmx_walltime_accounting_t walltime_accounting = nullptr;
754 MembedHolder membedHolder(filenames.size(), filenames.data());
756 /* CAUTION: threads may be started later on in this function, so
757 cr doesn't reflect the final parallel state right now */
760 /* TODO: inputrec should tell us whether we use an algorithm, not a file option */
761 const bool doEssentialDynamics = opt2bSet("-ei", filenames.size(), filenames.data());
762 const bool doRerun = mdrunOptions.rerun;
764 // Handle task-assignment related user options.
765 EmulateGpuNonbonded emulateGpuNonbonded =
766 (getenv("GMX_EMULATE_GPU") != nullptr ? EmulateGpuNonbonded::Yes : EmulateGpuNonbonded::No);
768 std::vector<int> userGpuTaskAssignment;
771 userGpuTaskAssignment = parseUserTaskAssignmentString(hw_opt.userGpuTaskAssignment);
773 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
774 auto nonbondedTarget = findTaskTarget(nbpu_opt);
775 auto pmeTarget = findTaskTarget(pme_opt);
776 auto pmeFftTarget = findTaskTarget(pme_fft_opt);
777 auto bondedTarget = findTaskTarget(bonded_opt);
778 auto updateTarget = findTaskTarget(update_opt);
780 FILE* fplog = nullptr;
781 // If we are appending, we don't write log output because we need
782 // to check that the old log file matches what the checkpoint file
783 // expects. Otherwise, we should start to write log output now if
784 // there is a file ready for it.
785 if (logFileHandle != nullptr && startingBehavior != StartingBehavior::RestartWithAppending)
787 fplog = gmx_fio_getfp(logFileHandle);
789 const bool isSimulationMasterRank = findIsSimulationMasterRank(ms, simulationCommunicator);
790 gmx::LoggerOwner logOwner(buildLogger(fplog, isSimulationMasterRank));
791 gmx::MDLogger mdlog(logOwner.logger());
793 gmx_print_detected_hardware(fplog, isSimulationMasterRank && isMasterSim(ms), mdlog, hwinfo_);
795 std::vector<int> gpuIdsToUse = makeGpuIdsToUse(hwinfo_->deviceInfoList, hw_opt.gpuIdsAvailable);
796 const int numDevicesToUse = gmx::ssize(gpuIdsToUse);
798 // Print citation requests after all software/hardware printing
799 pleaseCiteGromacs(fplog);
801 // Note: legacy program logic relies on checking whether these pointers are assigned.
802 // Objects may or may not be allocated later.
803 std::unique_ptr<t_inputrec> inputrec;
804 std::unique_ptr<t_state> globalState;
806 auto partialDeserializedTpr = std::make_unique<PartialDeserializedTprFile>();
808 if (isSimulationMasterRank)
810 // Allocate objects to be initialized by later function calls.
811 /* Only the master rank has the global state */
812 globalState = std::make_unique<t_state>();
813 inputrec = std::make_unique<t_inputrec>();
815 /* Read (nearly) all data required for the simulation
816 * and keep the partly serialized tpr contents to send to other ranks later
818 applyGlobalSimulationState(
819 *inputHolder_.get(), partialDeserializedTpr.get(), globalState.get(), inputrec.get(), &mtop);
822 /* Check and update the hardware options for internal consistency */
823 checkAndUpdateHardwareOptions(
824 mdlog, &hw_opt, isSimulationMasterRank, domdecOptions.numPmeRanks, inputrec.get());
826 if (GMX_THREAD_MPI && isSimulationMasterRank)
828 bool useGpuForNonbonded = false;
829 bool useGpuForPme = false;
832 GMX_RELEASE_ASSERT(inputrec != nullptr, "Keep the compiler happy");
834 // If the user specified the number of ranks, then we must
835 // respect that, but in default mode, we need to allow for
836 // the number of GPUs to choose the number of ranks.
837 auto canUseGpuForNonbonded = buildSupportsNonbondedOnGpu(nullptr);
838 useGpuForNonbonded = decideWhetherToUseGpusForNonbondedWithThreadMpi(
841 userGpuTaskAssignment,
843 canUseGpuForNonbonded,
844 gpuAccelerationOfNonbondedIsUseful(mdlog, *inputrec, GMX_THREAD_MPI),
845 hw_opt.nthreads_tmpi);
846 useGpuForPme = decideWhetherToUseGpusForPmeWithThreadMpi(useGpuForNonbonded,
849 userGpuTaskAssignment,
852 hw_opt.nthreads_tmpi,
853 domdecOptions.numPmeRanks);
855 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
857 /* Determine how many thread-MPI ranks to start.
859 * TODO Over-writing the user-supplied value here does
860 * prevent any possible subsequent checks from working
862 hw_opt.nthreads_tmpi = get_nthreads_mpi(hwinfo_,
870 membedHolder.doMembed());
872 // Now start the threads for thread MPI.
873 spawnThreads(hw_opt.nthreads_tmpi);
874 // The spawned threads enter mdrunner() and execution of
875 // master and spawned threads joins at the end of this block.
878 GMX_RELEASE_ASSERT(ms || simulationCommunicator != MPI_COMM_NULL,
879 "Must have valid communicator unless running a multi-simulation");
880 CommrecHandle crHandle = init_commrec(simulationCommunicator);
881 t_commrec* cr = crHandle.get();
882 GMX_RELEASE_ASSERT(cr != nullptr, "Must have valid commrec");
884 PhysicalNodeCommunicator physicalNodeComm(libraryWorldCommunicator, gmx_physicalnode_id_hash());
886 // If we detected the topology on this system, double-check that it makes sense
887 if (hwinfo_->hardwareTopology->isThisSystem())
889 hardwareTopologyDoubleCheckDetection(mdlog, *hwinfo_->hardwareTopology);
894 /* now broadcast everything to the non-master nodes/threads: */
895 if (!isSimulationMasterRank)
897 // Until now, only the master rank has a non-null pointer.
898 // On non-master ranks, allocate the object that will receive data in the following call.
899 inputrec = std::make_unique<t_inputrec>();
901 init_parallel(cr->mpiDefaultCommunicator,
905 partialDeserializedTpr.get());
907 GMX_RELEASE_ASSERT(inputrec != nullptr, "All ranks should have a valid inputrec now");
908 partialDeserializedTpr.reset(nullptr);
911 !inputrec->useConstantAcceleration,
912 "Linear acceleration has been removed in GROMACS 2022, and was broken for many years "
913 "before that. Use GROMACS 4.5 or earlier if you need this feature.");
915 // Now the number of ranks is known to all ranks, and each knows
916 // the inputrec read by the master rank. The ranks can now all run
917 // the task-deciding functions and will agree on the result
918 // without needing to communicate.
919 const bool useDomainDecomposition =
920 (PAR(cr) && !(EI_TPI(inputrec->eI) || inputrec->eI == IntegrationAlgorithm::NM));
922 // Note that these variables describe only their own node.
924 // Note that when bonded interactions run on a GPU they always run
925 // alongside a nonbonded task, so do not influence task assignment
926 // even though they affect the force calculation workload.
927 bool useGpuForNonbonded = false;
928 bool useGpuForPme = false;
929 bool useGpuForBonded = false;
930 bool useGpuForUpdate = false;
931 bool gpusWereDetected = hwinfo_->ngpu_compatible_tot > 0;
934 // It's possible that there are different numbers of GPUs on
935 // different nodes, which is the user's responsibility to
936 // handle. If unsuitable, we will notice that during task
938 auto canUseGpuForNonbonded = buildSupportsNonbondedOnGpu(nullptr);
939 useGpuForNonbonded = decideWhetherToUseGpusForNonbonded(
941 userGpuTaskAssignment,
943 canUseGpuForNonbonded,
944 gpuAccelerationOfNonbondedIsUseful(mdlog, *inputrec, !GMX_THREAD_MPI),
946 useGpuForPme = decideWhetherToUseGpusForPme(useGpuForNonbonded,
948 userGpuTaskAssignment,
951 cr->sizeOfDefaultCommunicator,
952 domdecOptions.numPmeRanks,
954 useGpuForBonded = decideWhetherToUseGpusForBonded(
955 useGpuForNonbonded, useGpuForPme, bondedTarget, *inputrec, mtop, domdecOptions.numPmeRanks, gpusWereDetected);
957 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
959 const PmeRunMode pmeRunMode = determinePmeRunMode(useGpuForPme, pmeFftTarget, *inputrec);
961 // Initialize development feature flags that enabled by environment variable
962 // and report those features that are enabled.
963 const DevelopmentFeatureFlags devFlags =
964 manageDevelopmentFeatures(mdlog, useGpuForNonbonded, pmeRunMode);
966 const bool useModularSimulator = checkUseModularSimulator(false,
974 membedHolder.doMembed());
977 // TODO: hide restraint implementation details from Mdrunner.
978 // There is nothing unique about restraints at this point as far as the
979 // Mdrunner is concerned. The Mdrunner should just be getting a sequence of
980 // factory functions from the SimulationContext on which to call mdModules_->add().
981 // TODO: capture all restraints into a single RestraintModule, passed to the runner builder.
982 for (auto&& restraint : restraintManager_->getRestraints())
984 auto module = RestraintMDModule::create(restraint, restraint->sites());
985 mdModules_->add(std::move(module));
988 // TODO: Error handling
989 mdModules_->assignOptionsToModules(*inputrec->params, nullptr);
990 // now that the MdModules know their options, they know which callbacks to sign up to
991 mdModules_->subscribeToSimulationSetupNotifications();
992 const auto& mdModulesNotifier = mdModules_->notifier().simulationSetupNotifications_;
994 if (inputrec->internalParameters != nullptr)
996 mdModulesNotifier.notify(*inputrec->internalParameters);
999 if (fplog != nullptr)
1001 pr_inputrec(fplog, 0, "Input Parameters", inputrec.get(), FALSE);
1002 fprintf(fplog, "\n");
1007 /* In rerun, set velocities to zero if present */
1008 if (doRerun && ((globalState->flags & enumValueToBitMask(StateEntry::V)) != 0))
1010 // rerun does not use velocities
1014 "Rerun trajectory contains velocities. Rerun does only evaluate "
1015 "potential energy and forces. The velocities will be ignored.");
1016 for (int i = 0; i < globalState->natoms; i++)
1018 clear_rvec(globalState->v[i]);
1020 globalState->flags &= ~enumValueToBitMask(StateEntry::V);
1023 /* now make sure the state is initialized and propagated */
1024 set_state_entries(globalState.get(), inputrec.get(), useModularSimulator);
1027 /* NM and TPI parallelize over force/energy calculations, not atoms,
1028 * so we need to initialize and broadcast the global state.
1030 if (inputrec->eI == IntegrationAlgorithm::NM || inputrec->eI == IntegrationAlgorithm::TPI)
1034 globalState = std::make_unique<t_state>();
1036 broadcastStateWithoutDynamics(
1037 cr->mpiDefaultCommunicator, DOMAINDECOMP(cr), PAR(cr), globalState.get());
1040 /* A parallel command line option consistency check that we can
1041 only do after any threads have started. */
1043 && (domdecOptions.numCells[XX] > 1 || domdecOptions.numCells[YY] > 1
1044 || domdecOptions.numCells[ZZ] > 1 || domdecOptions.numPmeRanks > 0))
1047 "The -dd or -npme option request a parallel simulation, "
1049 "but %s was compiled without threads or MPI enabled",
1050 output_env_get_program_display_name(oenv));
1051 #elif GMX_THREAD_MPI
1052 "but the number of MPI-threads (option -ntmpi) is not set or is 1");
1054 "but %s was not started through mpirun/mpiexec or only one rank was requested "
1055 "through mpirun/mpiexec",
1056 output_env_get_program_display_name(oenv));
1060 if (doRerun && (EI_ENERGY_MINIMIZATION(inputrec->eI) || IntegrationAlgorithm::NM == inputrec->eI))
1063 "The .mdp file specified an energy mininization or normal mode algorithm, and "
1064 "these are not compatible with mdrun -rerun");
1067 /* Object for collecting reasons for not using PME-only ranks */
1068 SeparatePmeRanksPermitted separatePmeRanksPermitted;
1070 /* Permit MDModules to notify whether they want to use PME-only ranks */
1071 mdModulesNotifier.notify(&separatePmeRanksPermitted);
1073 /* If simulation is not using PME then disable PME-only ranks */
1074 if (!(EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype)))
1076 separatePmeRanksPermitted.disablePmeRanks(
1077 "PME-only ranks are requested, but the system does not use PME "
1078 "for electrostatics or LJ");
1081 /* With NB GPUs we don't automatically use PME-only CPU ranks. PME ranks can
1082 * improve performance with many threads per GPU, since our OpenMP
1083 * scaling is bad, but it's difficult to automate the setup.
1085 if (useGpuForNonbonded && domdecOptions.numPmeRanks < 0)
1087 separatePmeRanksPermitted.disablePmeRanks(
1088 "PME-only CPU ranks are not automatically used when "
1089 "non-bonded interactions are computed on GPUs");
1092 /* If GPU is used for PME then only 1 PME rank is permitted */
1093 if (useGpuForPme && (domdecOptions.numPmeRanks < 0 || domdecOptions.numPmeRanks > 1))
1095 separatePmeRanksPermitted.disablePmeRanks(
1096 "PME GPU decomposition is not supported. Only one separate PME-only GPU rank "
1100 /* Disable PME-only ranks if some parts of the code requested so and it's up to GROMACS to decide */
1101 if (!separatePmeRanksPermitted.permitSeparatePmeRanks() && domdecOptions.numPmeRanks < 0)
1103 domdecOptions.numPmeRanks = 0;
1106 .appendText("Simulation will not use PME-only ranks because: "
1107 + separatePmeRanksPermitted.reasonsWhyDisabled());
1110 /* If some parts of the code could not use PME-only ranks and
1111 * user explicitly used mdrun -npme option then throw an error */
1112 if (!separatePmeRanksPermitted.permitSeparatePmeRanks() && domdecOptions.numPmeRanks > 0)
1114 gmx_fatal_collective(FARGS,
1115 cr->mpiDefaultCommunicator,
1117 "Requested -npme %d option is not viable because: %s",
1118 domdecOptions.numPmeRanks,
1119 separatePmeRanksPermitted.reasonsWhyDisabled().c_str());
1122 /* NMR restraints must be initialized before load_checkpoint,
1123 * since with time averaging the history is added to t_state.
1124 * For proper consistency check we therefore need to extend
1126 * So the PME-only nodes (if present) will also initialize
1127 * the distance restraints.
1130 /* This needs to be called before read_checkpoint to extend the state */
1131 t_disresdata* disresdata;
1132 snew(disresdata, 1);
1136 DisResRunMode::MDRun,
1137 MASTER(cr) ? DDRole::Master : DDRole::Agent,
1138 PAR(cr) ? NumRanks::Multiple : NumRanks::Single,
1143 replExParams.exchangeInterval > 0);
1145 t_oriresdata* oriresdata;
1146 snew(oriresdata, 1);
1147 init_orires(fplog, &mtop, inputrec.get(), cr, ms, globalState.get(), oriresdata);
1149 auto deform = prepareBoxDeformation(globalState != nullptr ? globalState->box : box,
1150 MASTER(cr) ? DDRole::Master : DDRole::Agent,
1151 PAR(cr) ? NumRanks::Multiple : NumRanks::Single,
1152 cr->mpi_comm_mygroup,
1156 /* We have to remember the generation's first step before reading checkpoint.
1157 This way, we can report to the F@H core both the generation's first step
1158 and the restored first step, thus making it able to distinguish between
1159 an interruption/resume and start of the n-th generation simulation.
1160 Having this information, the F@H core can correctly calculate and report
1163 int gen_first_step = 0;
1166 gen_first_step = inputrec->init_step;
1170 ObservablesHistory observablesHistory = {};
1172 auto modularSimulatorCheckpointData = std::make_unique<ReadCheckpointDataHolder>();
1173 if (startingBehavior != StartingBehavior::NewSimulation)
1175 /* Check if checkpoint file exists before doing continuation.
1176 * This way we can use identical input options for the first and subsequent runs...
1178 if (mdrunOptions.numStepsCommandline > -2)
1180 /* Temporarily set the number of steps to unlimited to avoid
1181 * triggering the nsteps check in load_checkpoint().
1182 * This hack will go away soon when the -nsteps option is removed.
1184 inputrec->nsteps = -1;
1187 // Finish applying initial simulation state information from external sources on all ranks.
1188 // Reconcile checkpoint file data with Mdrunner state established up to this point.
1189 applyLocalState(*inputHolder_.get(),
1192 domdecOptions.numCells,
1195 &observablesHistory,
1196 mdrunOptions.reproducible,
1197 mdModules_->notifier(),
1198 modularSimulatorCheckpointData.get(),
1199 useModularSimulator);
1200 // TODO: (#3652) Synchronize filesystem state, SimulationInput contents, and program
1202 // on all code paths.
1203 // Write checkpoint or provide hook to update SimulationInput.
1204 // If there was a checkpoint file, SimulationInput contains more information
1205 // than if there wasn't. At this point, we have synchronized the in-memory
1206 // state with the filesystem state only for restarted simulations. We should
1207 // be calling applyLocalState unconditionally and expect that the completeness
1208 // of SimulationInput is not dependent on its creation method.
1210 if (startingBehavior == StartingBehavior::RestartWithAppending && logFileHandle)
1212 // Now we can start normal logging to the truncated log file.
1213 fplog = gmx_fio_getfp(logFileHandle);
1214 prepareLogAppending(fplog);
1215 logOwner = buildLogger(fplog, MASTER(cr));
1216 mdlog = logOwner.logger();
1223 fcRegisterSteps(inputrec->nsteps + inputrec->init_step, gen_first_step);
1227 if (mdrunOptions.numStepsCommandline > -2)
1232 "The -nsteps functionality is deprecated, and may be removed in a future "
1234 "Consider using gmx convert-tpr -nsteps or changing the appropriate .mdp "
1237 /* override nsteps with value set on the commandline */
1238 override_nsteps_cmdline(mdlog, mdrunOptions.numStepsCommandline, inputrec.get());
1240 if (isSimulationMasterRank)
1242 copy_mat(globalState->box, box);
1247 gmx_bcast(sizeof(box), box, cr->mpiDefaultCommunicator);
1250 if (inputrec->cutoff_scheme != CutoffScheme::Verlet)
1253 "This group-scheme .tpr file can no longer be run by mdrun. Please update to the "
1254 "Verlet scheme, or use an earlier version of GROMACS if necessary.");
1256 /* Update rlist and nstlist. */
1257 /* Note: prepare_verlet_scheme is calling increaseNstlist(...), which (while attempting to
1258 * increase rlist) tries to check if the newly chosen value fits with the DD scheme. As this is
1259 * run before any DD scheme is set up, this check is never executed. See #3334 for more details.
1261 prepare_verlet_scheme(fplog,
1267 useGpuForNonbonded || (emulateGpuNonbonded == EmulateGpuNonbonded::Yes),
1270 // This builder is necessary while we have multi-part construction
1271 // of DD. Before DD is constructed, we use the existence of
1272 // the builder object to indicate that further construction of DD
1274 std::unique_ptr<DomainDecompositionBuilder> ddBuilder;
1275 if (useDomainDecomposition)
1277 ddBuilder = std::make_unique<DomainDecompositionBuilder>(
1285 positionsFromStatePointer(globalState.get()));
1289 /* PME, if used, is done on all nodes with 1D decomposition */
1290 cr->nnodes = cr->sizeOfDefaultCommunicator;
1291 cr->sim_nodeid = cr->rankInDefaultCommunicator;
1292 cr->nodeid = cr->rankInDefaultCommunicator;
1294 cr->duty = (DUTY_PP | DUTY_PME);
1296 if (inputrec->pbcType == PbcType::Screw)
1298 gmx_fatal(FARGS, "pbc=screw is only implemented with domain decomposition");
1302 // Produce the task assignment for this rank - done after DD is constructed
1303 GpuTaskAssignments gpuTaskAssignments = GpuTaskAssignmentsBuilder::build(
1305 userGpuTaskAssignment,
1307 simulationCommunicator,
1315 thisRankHasDuty(cr, DUTY_PP),
1316 // TODO cr->duty & DUTY_PME should imply that a PME
1317 // algorithm is active, but currently does not.
1318 EEL_PME(inputrec->coulombtype) && thisRankHasDuty(cr, DUTY_PME));
1320 // Get the device handles for the modules, nullptr when no task is assigned.
1322 DeviceInformation* deviceInfo = gpuTaskAssignments.initDevice(&deviceId);
1324 // timing enabling - TODO put this in gpu_utils (even though generally this is just option handling?)
1325 bool useTiming = true;
1329 /* WARNING: CUDA timings are incorrect with multiple streams.
1330 * This is the main reason why they are disabled by default.
1332 // TODO: Consider turning on by default when we can detect nr of streams.
1333 useTiming = (getenv("GMX_ENABLE_GPU_TIMING") != nullptr);
1335 else if (GMX_GPU_OPENCL)
1337 useTiming = (getenv("GMX_DISABLE_GPU_TIMING") == nullptr);
1340 // TODO Currently this is always built, yet DD partition code
1341 // checks if it is built before using it. Probably it should
1342 // become an MDModule that is made only when another module
1343 // requires it (e.g. pull, CompEl, density fitting), so that we
1344 // don't update the local atom sets unilaterally every step.
1345 LocalAtomSetManager atomSets;
1348 // TODO Pass the GPU streams to ddBuilder to use in buffer
1349 // transfers (e.g. halo exchange)
1350 cr->dd = ddBuilder->build(&atomSets);
1351 // The builder's job is done, so destruct it
1352 ddBuilder.reset(nullptr);
1353 // Note that local state still does not exist yet.
1356 // The GPU update is decided here because we need to know whether the constraints or
1357 // SETTLEs can span accross the domain borders (i.e. whether or not update groups are
1358 // defined). This is only known after DD is initialized, hence decision on using GPU
1359 // update is done so late.
1362 const bool useUpdateGroups = cr->dd ? ddUsesUpdateGroups(*cr->dd) : false;
1364 useGpuForUpdate = decideWhetherToUseGpuForUpdate(useDomainDecomposition,
1367 domdecOptions.numPmeRanks > 0,
1373 doEssentialDynamics,
1374 gmx_mtop_ftype_count(mtop, F_ORIRES) > 0,
1375 replExParams.exchangeInterval > 0,
1376 inputrec->opts.nFreeze != nullptr,
1381 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1383 const bool printHostName = (cr->nnodes > 1);
1384 gpuTaskAssignments.reportGpuUsage(mdlog, printHostName, useGpuForBonded, pmeRunMode, useGpuForUpdate);
1386 const bool disableNonbondedCalculation = (getenv("GMX_NO_NONBONDED") != nullptr);
1387 if (disableNonbondedCalculation)
1389 /* turn off non-bonded calculations */
1390 GMX_LOG(mdlog.warning)
1393 "Found environment variable GMX_NO_NONBONDED.\n"
1394 "Disabling nonbonded calculations.");
1397 MdrunScheduleWorkload runScheduleWork;
1399 bool useGpuDirectHalo = decideWhetherToUseGpuForHalo(devFlags,
1400 havePPDomainDecomposition(cr),
1402 useModularSimulator,
1404 EI_ENERGY_MINIMIZATION(inputrec->eI));
1406 // Also populates the simulation constant workload description.
1407 runScheduleWork.simulationWork = createSimulationWorkload(*inputrec,
1408 disableNonbondedCalculation,
1416 std::unique_ptr<DeviceStreamManager> deviceStreamManager = nullptr;
1418 if (deviceInfo != nullptr)
1420 if (DOMAINDECOMP(cr) && thisRankHasDuty(cr, DUTY_PP))
1422 dd_setup_dlb_resource_sharing(cr, deviceId);
1424 deviceStreamManager = std::make_unique<DeviceStreamManager>(
1425 *deviceInfo, havePPDomainDecomposition(cr), runScheduleWork.simulationWork, useTiming);
1428 // If the user chose a task assignment, give them some hints
1429 // where appropriate.
1430 if (!userGpuTaskAssignment.empty())
1432 gpuTaskAssignments.logPerformanceHints(mdlog, numDevicesToUse);
1437 /* After possible communicator splitting in make_dd_communicators.
1438 * we can set up the intra/inter node communication.
1440 gmx_setup_nodecomm(fplog, cr);
1446 GMX_LOG(mdlog.warning)
1448 .appendTextFormatted(
1449 "This is simulation %d out of %d running as a composite GROMACS\n"
1450 "multi-simulation job. Setup for this simulation:\n",
1451 ms->simulationIndex_,
1452 ms->numSimulations_);
1454 GMX_LOG(mdlog.warning)
1455 .appendTextFormatted("Using %d MPI %s\n",
1458 cr->nnodes == 1 ? "thread" : "threads"
1460 cr->nnodes == 1 ? "process" : "processes"
1466 // If mdrun -pin auto honors any affinity setting that already
1467 // exists. If so, it is nice to provide feedback about whether
1468 // that existing affinity setting was from OpenMP or something
1469 // else, so we run this code both before and after we initialize
1470 // the OpenMP support.
1471 gmx_check_thread_affinity_set(mdlog, &hw_opt, hwinfo_->nthreads_hw_avail, FALSE);
1472 /* Check and update the number of OpenMP threads requested */
1473 checkAndUpdateRequestedNumOpenmpThreads(
1474 &hw_opt, *hwinfo_, cr, ms, physicalNodeComm.size_, pmeRunMode, mtop, *inputrec);
1476 gmx_omp_nthreads_init(mdlog,
1478 hwinfo_->nthreads_hw_avail,
1479 physicalNodeComm.size_,
1480 hw_opt.nthreads_omp,
1481 hw_opt.nthreads_omp_pme,
1482 !thisRankHasDuty(cr, DUTY_PP));
1484 const bool bEnableFPE = gmxShouldEnableFPExceptions();
1485 // FIXME - reconcile with gmx_feenableexcept() call from CommandLineModuleManager::run()
1488 gmx_feenableexcept();
1491 /* Now that we know the setup is consistent, check for efficiency */
1492 check_resource_division_efficiency(
1493 hwinfo_, gpuTaskAssignments.thisRankHasAnyGpuTask(), mdrunOptions.ntompOptionIsSet, cr, mdlog);
1495 /* getting number of PP/PME threads on this MPI / tMPI rank.
1496 PME: env variable should be read only on one node to make sure it is
1497 identical everywhere;
1499 const int numThreadsOnThisRank = thisRankHasDuty(cr, DUTY_PP) ? gmx_omp_nthreads_get(emntNonbonded)
1500 : gmx_omp_nthreads_get(emntPME);
1501 checkHardwareOversubscription(
1502 numThreadsOnThisRank, cr->nodeid, *hwinfo_->hardwareTopology, physicalNodeComm, mdlog);
1504 // Enable Peer access between GPUs where available
1505 // Only for DD, only master PP rank needs to perform setup, and only if thread MPI plus
1506 // any of the GPU communication features are active.
1507 if (DOMAINDECOMP(cr) && MASTER(cr) && thisRankHasDuty(cr, DUTY_PP) && GMX_THREAD_MPI
1508 && (runScheduleWork.simulationWork.useGpuHaloExchange
1509 || runScheduleWork.simulationWork.useGpuPmePpCommunication))
1511 setupGpuDevicePeerAccess(gpuIdsToUse, mdlog);
1514 if (hw_opt.threadAffinity != ThreadAffinity::Off)
1516 /* Before setting affinity, check whether the affinity has changed
1517 * - which indicates that probably the OpenMP library has changed it
1518 * since we first checked).
1520 gmx_check_thread_affinity_set(mdlog, &hw_opt, hwinfo_->nthreads_hw_avail, TRUE);
1522 int numThreadsOnThisNode, intraNodeThreadOffset;
1523 analyzeThreadsOnThisNode(
1524 physicalNodeComm, numThreadsOnThisRank, &numThreadsOnThisNode, &intraNodeThreadOffset);
1526 /* Set the CPU affinity */
1527 gmx_set_thread_affinity(mdlog,
1530 *hwinfo_->hardwareTopology,
1531 numThreadsOnThisRank,
1532 numThreadsOnThisNode,
1533 intraNodeThreadOffset,
1537 if (mdrunOptions.timingOptions.resetStep > -1)
1542 "The -resetstep functionality is deprecated, and may be removed in a "
1545 wcycle = wallcycle_init(fplog, mdrunOptions.timingOptions.resetStep, cr);
1549 /* Master synchronizes its value of reset_counters with all nodes
1550 * including PME only nodes */
1551 int64_t reset_counters = wcycle_get_reset_counters(wcycle);
1552 gmx_bcast(sizeof(reset_counters), &reset_counters, cr->mpi_comm_mysim);
1553 wcycle_set_reset_counters(wcycle, reset_counters);
1556 // Membrane embedding must be initialized before we call init_forcerec()
1557 membedHolder.initializeMembed(fplog,
1564 &mdrunOptions.checkpointOptions.period);
1566 const bool thisRankHasPmeGpuTask = gpuTaskAssignments.thisRankHasPmeGpuTask();
1567 std::unique_ptr<MDAtoms> mdAtoms;
1568 std::unique_ptr<VirtualSitesHandler> vsite;
1569 std::unique_ptr<GpuBonded> gpuBonded;
1572 if (thisRankHasDuty(cr, DUTY_PP))
1574 mdModulesNotifier.notify(*cr);
1575 mdModulesNotifier.notify(&atomSets);
1576 mdModulesNotifier.notify(mtop);
1577 mdModulesNotifier.notify(inputrec->pbcType);
1578 mdModulesNotifier.notify(SimulationTimeStep{ inputrec->delta_t });
1579 /* Initiate forcerecord */
1580 fr = std::make_unique<t_forcerec>();
1581 fr->forceProviders = mdModules_->initForceProviders();
1582 init_forcerec(fplog,
1589 opt2fn("-table", filenames.size(), filenames.data()),
1590 opt2fn("-tablep", filenames.size(), filenames.data()),
1591 opt2fns("-tableb", filenames.size(), filenames.data()),
1593 // Dirty hack, for fixing disres and orires should be made mdmodules
1594 fr->fcdata->disres = disresdata;
1595 fr->fcdata->orires = oriresdata;
1597 // Save a handle to device stream manager to use elsewhere in the code
1598 // TODO: Forcerec is not a correct place to store it.
1599 fr->deviceStreamManager = deviceStreamManager.get();
1601 if (runScheduleWork.simulationWork.useGpuPmePpCommunication && !thisRankHasDuty(cr, DUTY_PME))
1604 deviceStreamManager != nullptr,
1605 "GPU device stream manager should be valid in order to use PME-PP direct "
1608 deviceStreamManager->streamIsValid(DeviceStreamType::PmePpTransfer),
1609 "GPU PP-PME stream should be valid in order to use GPU PME-PP direct "
1611 fr->pmePpCommGpu = std::make_unique<gmx::PmePpCommGpu>(
1614 deviceStreamManager->context(),
1615 deviceStreamManager->stream(DeviceStreamType::PmePpTransfer));
1618 fr->nbv = Nbnxm::init_nb_verlet(mdlog,
1623 runScheduleWork.simulationWork.useGpuNonbonded,
1624 deviceStreamManager.get(),
1628 // TODO: Move the logic below to a GPU bonded builder
1629 if (runScheduleWork.simulationWork.useGpuBonded)
1631 GMX_RELEASE_ASSERT(deviceStreamManager != nullptr,
1632 "GPU device stream manager should be valid in order to use GPU "
1633 "version of bonded forces.");
1634 gpuBonded = std::make_unique<GpuBonded>(
1636 fr->ic->epsfac * fr->fudgeQQ,
1637 deviceStreamManager->context(),
1638 deviceStreamManager->bondedStream(havePPDomainDecomposition(cr)),
1640 fr->gpuBonded = gpuBonded.get();
1643 /* Initialize the mdAtoms structure.
1644 * mdAtoms is not filled with atom data,
1645 * as this can not be done now with domain decomposition.
1647 mdAtoms = makeMDAtoms(fplog, mtop, *inputrec, thisRankHasPmeGpuTask);
1648 if (globalState && thisRankHasPmeGpuTask)
1650 // The pinning of coordinates in the global state object works, because we only use
1651 // PME on GPU without DD or on a separate PME rank, and because the local state pointer
1652 // points to the global state object without DD.
1653 // FIXME: MD and EM separately set up the local state - this should happen in the same
1654 // function, which should also perform the pinning.
1655 changePinningPolicy(&globalState->x, pme_get_pinning_policy());
1658 /* Initialize the virtual site communication */
1659 vsite = makeVirtualSitesHandler(mtop, cr, fr->pbcType);
1661 calc_shifts(box, fr->shift_vec);
1663 /* With periodic molecules the charge groups should be whole at start up
1664 * and the virtual sites should not be far from their proper positions.
1666 if (!inputrec->bContinuation && MASTER(cr)
1667 && !(inputrec->pbcType != PbcType::No && inputrec->bPeriodicMols))
1669 /* Make molecules whole at start of run */
1670 if (fr->pbcType != PbcType::No)
1672 do_pbc_first_mtop(fplog, inputrec->pbcType, box, &mtop, globalState->x.rvec_array());
1676 /* Correct initial vsite positions are required
1677 * for the initial distribution in the domain decomposition
1678 * and for the initial shell prediction.
1680 constructVirtualSitesGlobal(mtop, globalState->x);
1684 if (EEL_PME(fr->ic->eeltype) || EVDW_PME(fr->ic->vdwtype))
1686 ewaldcoeff_q = fr->ic->ewaldcoeff_q;
1687 ewaldcoeff_lj = fr->ic->ewaldcoeff_lj;
1692 /* This is a PME only node */
1694 GMX_ASSERT(globalState == nullptr,
1695 "We don't need the state on a PME only rank and expect it to be unitialized");
1697 ewaldcoeff_q = calc_ewaldcoeff_q(inputrec->rcoulomb, inputrec->ewald_rtol);
1698 ewaldcoeff_lj = calc_ewaldcoeff_lj(inputrec->rvdw, inputrec->ewald_rtol_lj);
1701 gmx_pme_t* sepPmeData = nullptr;
1702 // This reference hides the fact that PME data is owned by runner on PME-only ranks and by forcerec on other ranks
1703 GMX_ASSERT(thisRankHasDuty(cr, DUTY_PP) == (fr != nullptr),
1704 "Double-checking that only PME-only ranks have no forcerec");
1705 gmx_pme_t*& pmedata = fr ? fr->pmedata : sepPmeData;
1707 // TODO should live in ewald module once its testing is improved
1709 // Later, this program could contain kernels that might be later
1710 // re-used as auto-tuning progresses, or subsequent simulations
1712 PmeGpuProgramStorage pmeGpuProgram;
1713 if (thisRankHasPmeGpuTask)
1716 (deviceStreamManager != nullptr),
1717 "GPU device stream manager should be initialized in order to use GPU for PME.");
1718 GMX_RELEASE_ASSERT((deviceInfo != nullptr),
1719 "GPU device should be initialized in order to use GPU for PME.");
1720 pmeGpuProgram = buildPmeGpuProgram(deviceStreamManager->context());
1723 /* Initiate PME if necessary,
1724 * either on all nodes or on dedicated PME nodes only. */
1725 if (EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype))
1727 if (mdAtoms && mdAtoms->mdatoms())
1729 nChargePerturbed = mdAtoms->mdatoms()->nChargePerturbed;
1730 if (EVDW_PME(inputrec->vdwtype))
1732 nTypePerturbed = mdAtoms->mdatoms()->nTypePerturbed;
1735 if (cr->npmenodes > 0)
1737 /* The PME only nodes need to know nChargePerturbed(FEP on Q) and nTypePerturbed(FEP on LJ)*/
1738 gmx_bcast(sizeof(nChargePerturbed), &nChargePerturbed, cr->mpi_comm_mysim);
1739 gmx_bcast(sizeof(nTypePerturbed), &nTypePerturbed, cr->mpi_comm_mysim);
1742 if (thisRankHasDuty(cr, DUTY_PME))
1746 // TODO: This should be in the builder.
1747 GMX_RELEASE_ASSERT(!runScheduleWork.simulationWork.useGpuPme
1748 || (deviceStreamManager != nullptr),
1749 "Device stream manager should be valid in order to use GPU "
1752 !runScheduleWork.simulationWork.useGpuPme
1753 || deviceStreamManager->streamIsValid(DeviceStreamType::Pme),
1754 "GPU PME stream should be valid in order to use GPU version of PME.");
1756 const DeviceContext* deviceContext = runScheduleWork.simulationWork.useGpuPme
1757 ? &deviceStreamManager->context()
1759 const DeviceStream* pmeStream =
1760 runScheduleWork.simulationWork.useGpuPme
1761 ? &deviceStreamManager->stream(DeviceStreamType::Pme)
1764 pmedata = gmx_pme_init(cr,
1765 getNumPmeDomains(cr->dd),
1767 nChargePerturbed != 0,
1768 nTypePerturbed != 0,
1769 mdrunOptions.reproducible,
1772 gmx_omp_nthreads_get(emntPME),
1777 pmeGpuProgram.get(),
1780 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1785 if (EI_DYNAMICS(inputrec->eI))
1787 /* Turn on signal handling on all nodes */
1789 * (A user signal from the PME nodes (if any)
1790 * is communicated to the PP nodes.
1792 signal_handler_install();
1795 pull_t* pull_work = nullptr;
1796 if (thisRankHasDuty(cr, DUTY_PP))
1798 /* Assumes uniform use of the number of OpenMP threads */
1799 walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(emntDefault));
1801 if (inputrec->bPull)
1803 /* Initialize pull code */
1804 pull_work = init_pull(fplog,
1805 inputrec->pull.get(),
1810 inputrec->fepvals->init_lambda);
1811 if (inputrec->pull->bXOutAverage || inputrec->pull->bFOutAverage)
1813 initPullHistory(pull_work, &observablesHistory);
1815 if (EI_DYNAMICS(inputrec->eI) && MASTER(cr))
1817 init_pull_output_files(pull_work, filenames.size(), filenames.data(), oenv, startingBehavior);
1821 std::unique_ptr<EnforcedRotation> enforcedRotation;
1824 /* Initialize enforced rotation code */
1825 enforcedRotation = init_rot(fplog,
1838 t_swap* swap = nullptr;
1839 if (inputrec->eSwapCoords != SwapType::No)
1841 /* Initialize ion swapping code */
1842 swap = init_swapcoords(fplog,
1844 opt2fn_master("-swap", filenames.size(), filenames.data(), cr),
1847 &observablesHistory,
1855 /* Let makeConstraints know whether we have essential dynamics constraints. */
1856 auto constr = makeConstraints(
1857 mtop, *inputrec, pull_work, doEssentialDynamics, fplog, cr, ms, &nrnb, wcycle, fr->bMolPBC);
1859 /* Energy terms and groups */
1860 gmx_enerdata_t enerd(mtop.groups.groups[SimulationAtomGroupType::EnergyOutput].size(),
1861 inputrec->fepvals->n_lambda);
1863 // cos acceleration is only supported by md, but older tpr
1864 // files might still combine it with other integrators
1865 GMX_RELEASE_ASSERT(inputrec->cos_accel == 0.0 || inputrec->eI == IntegrationAlgorithm::MD,
1866 "cos_acceleration is only supported by integrator=md");
1868 /* Kinetic energy data */
1869 gmx_ekindata_t ekind;
1870 init_ekindata(fplog, &(inputrec->opts), &ekind, inputrec->cos_accel);
1872 /* Set up interactive MD (IMD) */
1873 auto imdSession = makeImdSession(inputrec.get(),
1880 MASTER(cr) ? globalState->x.rvec_array() : nullptr,
1884 mdrunOptions.imdOptions,
1887 if (DOMAINDECOMP(cr))
1889 GMX_RELEASE_ASSERT(fr, "fr was NULL while cr->duty was DUTY_PP");
1890 /* This call is not included in init_domain_decomposition mainly
1891 * because fr->cginfo_mb is set later.
1893 dd_init_bondeds(fplog,
1898 domdecOptions.checkBondedInteractions ? DDBondedChecking::All
1899 : DDBondedChecking::ExcludeZeroLimit,
1903 if (runScheduleWork.simulationWork.useGpuBufferOps)
1905 fr->gpuForceReduction[gmx::AtomLocality::Local] = std::make_unique<gmx::GpuForceReduction>(
1906 deviceStreamManager->context(),
1907 deviceStreamManager->stream(gmx::DeviceStreamType::NonBondedLocal),
1909 fr->gpuForceReduction[gmx::AtomLocality::NonLocal] = std::make_unique<gmx::GpuForceReduction>(
1910 deviceStreamManager->context(),
1911 deviceStreamManager->stream(gmx::DeviceStreamType::NonBondedNonLocal),
1915 std::unique_ptr<gmx::StatePropagatorDataGpu> stateGpu;
1916 if (gpusWereDetected
1917 && ((runScheduleWork.simulationWork.useGpuPme && thisRankHasDuty(cr, DUTY_PME))
1918 || runScheduleWork.simulationWork.useGpuBufferOps))
1920 GpuApiCallBehavior transferKind =
1921 (inputrec->eI == IntegrationAlgorithm::MD && !doRerun && !useModularSimulator)
1922 ? GpuApiCallBehavior::Async
1923 : GpuApiCallBehavior::Sync;
1924 GMX_RELEASE_ASSERT(deviceStreamManager != nullptr,
1925 "GPU device stream manager should be initialized to use GPU.");
1926 stateGpu = std::make_unique<gmx::StatePropagatorDataGpu>(
1927 *deviceStreamManager, transferKind, pme_gpu_get_block_size(fr->pmedata), wcycle);
1928 fr->stateGpu = stateGpu.get();
1931 GMX_ASSERT(stopHandlerBuilder_, "Runner must provide StopHandlerBuilder to simulator.");
1932 SimulatorBuilder simulatorBuilder;
1934 simulatorBuilder.add(SimulatorStateData(globalState.get(), &observablesHistory, &enerd, &ekind));
1935 simulatorBuilder.add(std::move(membedHolder));
1936 simulatorBuilder.add(std::move(stopHandlerBuilder_));
1937 simulatorBuilder.add(SimulatorConfig(mdrunOptions, startingBehavior, &runScheduleWork));
1940 simulatorBuilder.add(SimulatorEnv(fplog, cr, ms, mdlog, oenv));
1941 simulatorBuilder.add(Profiling(&nrnb, walltime_accounting, wcycle));
1942 simulatorBuilder.add(ConstraintsParam(
1943 constr.get(), enforcedRotation ? enforcedRotation->getLegacyEnfrot() : nullptr, vsite.get()));
1944 // TODO: Separate `fr` to a separate add, and make the `build` handle the coupling sensibly.
1945 simulatorBuilder.add(LegacyInput(
1946 static_cast<int>(filenames.size()), filenames.data(), inputrec.get(), fr.get()));
1947 simulatorBuilder.add(ReplicaExchangeParameters(replExParams));
1948 simulatorBuilder.add(InteractiveMD(imdSession.get()));
1949 simulatorBuilder.add(SimulatorModules(mdModules_->outputProvider(), mdModules_->notifier()));
1950 simulatorBuilder.add(CenterOfMassPulling(pull_work));
1951 // Todo move to an MDModule
1952 simulatorBuilder.add(IonSwapping(swap));
1953 simulatorBuilder.add(TopologyData(&mtop, mdAtoms.get()));
1954 simulatorBuilder.add(BoxDeformationHandle(deform.get()));
1955 simulatorBuilder.add(std::move(modularSimulatorCheckpointData));
1957 // build and run simulator object based on user-input
1958 auto simulator = simulatorBuilder.build(useModularSimulator);
1961 if (fr->pmePpCommGpu)
1963 // destroy object since it is no longer required. (This needs to be done while the GPU context still exists.)
1964 fr->pmePpCommGpu.reset();
1967 if (inputrec->bPull)
1969 finish_pull(pull_work);
1971 finish_swapcoords(swap);
1975 GMX_RELEASE_ASSERT(pmedata, "pmedata was NULL while cr->duty was not DUTY_PP");
1977 walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(emntPME));
1978 gmx_pmeonly(pmedata,
1982 walltime_accounting,
1985 deviceStreamManager.get());
1988 wallcycle_stop(wcycle, ewcRUN);
1990 /* Finish up, write some stuff
1991 * if rerunMD, don't write last frame again
1999 walltime_accounting,
2000 fr ? fr->nbv.get() : nullptr,
2002 EI_DYNAMICS(inputrec->eI) && !isMultiSim(ms));
2004 // clean up cycle counter
2005 wallcycle_destroy(wcycle);
2007 deviceStreamManager.reset(nullptr);
2011 gmx_pme_destroy(pmedata);
2015 // FIXME: this is only here to manually unpin mdAtoms->chargeA_ and state->x,
2016 // before we destroy the GPU context(s)
2017 // Pinned buffers are associated with contexts in CUDA.
2018 // As soon as we destroy GPU contexts after mdrunner() exits, these lines should go.
2019 mdAtoms.reset(nullptr);
2020 globalState.reset(nullptr);
2021 mdModules_.reset(nullptr); // destruct force providers here as they might also use the GPU
2022 gpuBonded.reset(nullptr);
2023 fr.reset(nullptr); // destruct forcerec before gpu
2024 // TODO convert to C++ so we can get rid of these frees
2028 if (!hwinfo_->deviceInfoList.empty())
2030 /* stop the GPU profiler (only CUDA) */
2034 /* With tMPI we need to wait for all ranks to finish deallocation before
2035 * destroying the CUDA context as some tMPI ranks may be sharing
2038 * This is not a concern in OpenCL where we use one context per rank.
2040 * Note: it is safe to not call the barrier on the ranks which do not use GPU,
2041 * but it is easier and more futureproof to call it on the whole node.
2043 * Note that this function needs to be called even if GPUs are not used
2044 * in this run because the PME ranks have no knowledge of whether GPUs
2045 * are used or not, but all ranks need to enter the barrier below.
2046 * \todo Remove this physical node barrier after making sure
2047 * that it's not needed anymore (with a shared GPU run).
2051 physicalNodeComm.barrier();
2053 releaseDevice(deviceInfo);
2055 /* Does what it says */
2056 print_date_and_time(fplog, cr->nodeid, "Finished mdrun", gmx_gettime());
2057 walltime_accounting_destroy(walltime_accounting);
2059 // Ensure log file content is written
2062 gmx_fio_flush(logFileHandle);
2065 /* Reset FPEs (important for unit tests) by disabling them. Assumes no
2066 * exceptions were enabled before function was called. */
2069 gmx_fedisableexcept();
2072 auto rc = static_cast<int>(gmx_get_stop_condition());
2075 /* we need to join all threads. The sub-threads join when they
2076 exit this function, but the master thread needs to be told to
2086 Mdrunner::~Mdrunner()
2088 // Clean up of the Manager.
2089 // This will end up getting called on every thread-MPI rank, which is unnecessary,
2090 // but okay as long as threads synchronize some time before adding or accessing
2091 // a new set of restraints.
2092 if (restraintManager_)
2094 restraintManager_->clear();
2095 GMX_ASSERT(restraintManager_->countRestraints() == 0,
2096 "restraints added during runner life time should be cleared at runner "
2101 void Mdrunner::addPotential(std::shared_ptr<gmx::IRestraintPotential> puller, const std::string& name)
2103 GMX_ASSERT(restraintManager_, "Mdrunner must have a restraint manager.");
2104 // Not sure if this should be logged through the md logger or something else,
2105 // but it is helpful to have some sort of INFO level message sent somewhere.
2106 // std::cout << "Registering restraint named " << name << std::endl;
2108 // When multiple restraints are used, it may be wasteful to register them separately.
2109 // Maybe instead register an entire Restraint Manager as a force provider.
2110 restraintManager_->addToSpec(std::move(puller), name);
2113 Mdrunner::Mdrunner(std::unique_ptr<MDModules> mdModules) : mdModules_(std::move(mdModules)) {}
2115 Mdrunner::Mdrunner(Mdrunner&&) noexcept = default;
2117 //NOLINTNEXTLINE(performance-noexcept-move-constructor) working around GCC bug 58265 in CentOS 7
2118 Mdrunner& Mdrunner::operator=(Mdrunner&& /*handle*/) noexcept(BUGFREE_NOEXCEPT_STRING) = default;
2120 class Mdrunner::BuilderImplementation
2123 BuilderImplementation() = delete;
2124 BuilderImplementation(std::unique_ptr<MDModules> mdModules, compat::not_null<SimulationContext*> context);
2125 ~BuilderImplementation();
2127 BuilderImplementation& setExtraMdrunOptions(const MdrunOptions& options,
2128 real forceWarningThreshold,
2129 StartingBehavior startingBehavior);
2131 void addHardwareDetectionResult(const gmx_hw_info_t* hwinfo);
2133 void addDomdec(const DomdecOptions& options);
2135 void addInput(SimulationInputHandle inputHolder);
2137 void addVerletList(int nstlist);
2139 void addReplicaExchange(const ReplicaExchangeParameters& params);
2141 void addNonBonded(const char* nbpu_opt);
2143 void addPME(const char* pme_opt_, const char* pme_fft_opt_);
2145 void addBondedTaskAssignment(const char* bonded_opt);
2147 void addUpdateTaskAssignment(const char* update_opt);
2149 void addHardwareOptions(const gmx_hw_opt_t& hardwareOptions);
2151 void addFilenames(ArrayRef<const t_filenm> filenames);
2153 void addOutputEnvironment(gmx_output_env_t* outputEnvironment);
2155 void addLogFile(t_fileio* logFileHandle);
2157 void addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder);
2162 // Default parameters copied from runner.h
2163 // \todo Clarify source(s) of default parameters.
2165 const char* nbpu_opt_ = nullptr;
2166 const char* pme_opt_ = nullptr;
2167 const char* pme_fft_opt_ = nullptr;
2168 const char* bonded_opt_ = nullptr;
2169 const char* update_opt_ = nullptr;
2171 MdrunOptions mdrunOptions_;
2173 DomdecOptions domdecOptions_;
2175 ReplicaExchangeParameters replicaExchangeParameters_;
2177 //! Command-line override for the duration of a neighbor list with the Verlet scheme.
2180 //! World communicator, used for hardware detection and task assignment
2181 MPI_Comm libraryWorldCommunicator_ = MPI_COMM_NULL;
2183 //! Multisim communicator handle.
2184 gmx_multisim_t* multiSimulation_;
2186 //! mdrun communicator
2187 MPI_Comm simulationCommunicator_ = MPI_COMM_NULL;
2189 //! Print a warning if any force is larger than this (in kJ/mol nm).
2190 real forceWarningThreshold_ = -1;
2192 //! Whether the simulation will start afresh, or restart with/without appending.
2193 StartingBehavior startingBehavior_ = StartingBehavior::NewSimulation;
2195 //! The modules that comprise the functionality of mdrun.
2196 std::unique_ptr<MDModules> mdModules_;
2198 //! Detected hardware.
2199 const gmx_hw_info_t* hwinfo_ = nullptr;
2201 //! \brief Parallelism information.
2202 gmx_hw_opt_t hardwareOptions_;
2204 //! filename options for simulation.
2205 ArrayRef<const t_filenm> filenames_;
2207 /*! \brief Handle to output environment.
2209 * \todo gmx_output_env_t needs lifetime management.
2211 gmx_output_env_t* outputEnvironment_ = nullptr;
2213 /*! \brief Non-owning handle to MD log file.
2215 * \todo Context should own output facilities for client.
2216 * \todo Improve log file handle management.
2218 * Code managing the FILE* relies on the ability to set it to
2219 * nullptr to check whether the filehandle is valid.
2221 t_fileio* logFileHandle_ = nullptr;
2224 * \brief Builder for simulation stop signal handler.
2226 std::unique_ptr<StopHandlerBuilder> stopHandlerBuilder_ = nullptr;
2229 * \brief Sources for initial simulation state.
2231 * See issue #3652 for near-term refinements to the SimulationInput interface.
2233 * See issue #3379 for broader discussion on API aspects of simulation inputs and outputs.
2235 SimulationInputHandle inputHolder_;
2238 Mdrunner::BuilderImplementation::BuilderImplementation(std::unique_ptr<MDModules> mdModules,
2239 compat::not_null<SimulationContext*> context) :
2240 mdModules_(std::move(mdModules))
2242 libraryWorldCommunicator_ = context->libraryWorldCommunicator_;
2243 simulationCommunicator_ = context->simulationCommunicator_;
2244 multiSimulation_ = context->multiSimulation_.get();
2247 Mdrunner::BuilderImplementation::~BuilderImplementation() = default;
2249 Mdrunner::BuilderImplementation&
2250 Mdrunner::BuilderImplementation::setExtraMdrunOptions(const MdrunOptions& options,
2251 const real forceWarningThreshold,
2252 const StartingBehavior startingBehavior)
2254 mdrunOptions_ = options;
2255 forceWarningThreshold_ = forceWarningThreshold;
2256 startingBehavior_ = startingBehavior;
2260 void Mdrunner::BuilderImplementation::addDomdec(const DomdecOptions& options)
2262 domdecOptions_ = options;
2265 void Mdrunner::BuilderImplementation::addVerletList(int nstlist)
2270 void Mdrunner::BuilderImplementation::addReplicaExchange(const ReplicaExchangeParameters& params)
2272 replicaExchangeParameters_ = params;
2275 Mdrunner Mdrunner::BuilderImplementation::build()
2277 auto newRunner = Mdrunner(std::move(mdModules_));
2279 newRunner.mdrunOptions = mdrunOptions_;
2280 newRunner.pforce = forceWarningThreshold_;
2281 newRunner.startingBehavior = startingBehavior_;
2282 newRunner.domdecOptions = domdecOptions_;
2284 // \todo determine an invariant to check or confirm that all gmx_hw_opt_t objects are valid
2285 newRunner.hw_opt = hardwareOptions_;
2287 // No invariant to check. This parameter exists to optionally override other behavior.
2288 newRunner.nstlist_cmdline = nstlist_;
2290 newRunner.replExParams = replicaExchangeParameters_;
2292 newRunner.filenames = filenames_;
2294 newRunner.libraryWorldCommunicator = libraryWorldCommunicator_;
2296 newRunner.simulationCommunicator = simulationCommunicator_;
2298 // nullptr is a valid value for the multisim handle
2299 newRunner.ms = multiSimulation_;
2303 newRunner.hwinfo_ = hwinfo_;
2307 GMX_THROW(gmx::APIError(
2308 "MdrunnerBuilder::addHardwareDetectionResult() is required before build()"));
2313 newRunner.inputHolder_ = std::move(inputHolder_);
2317 GMX_THROW(gmx::APIError("MdrunnerBuilder::addInput() is required before build()."));
2320 // \todo Clarify ownership and lifetime management for gmx_output_env_t
2321 // \todo Update sanity checking when output environment has clearly specified invariants.
2322 // Initialization and default values for oenv are not well specified in the current version.
2323 if (outputEnvironment_)
2325 newRunner.oenv = outputEnvironment_;
2329 GMX_THROW(gmx::APIError(
2330 "MdrunnerBuilder::addOutputEnvironment() is required before build()"));
2333 newRunner.logFileHandle = logFileHandle_;
2337 newRunner.nbpu_opt = nbpu_opt_;
2341 GMX_THROW(gmx::APIError("MdrunnerBuilder::addNonBonded() is required before build()"));
2344 if (pme_opt_ && pme_fft_opt_)
2346 newRunner.pme_opt = pme_opt_;
2347 newRunner.pme_fft_opt = pme_fft_opt_;
2351 GMX_THROW(gmx::APIError("MdrunnerBuilder::addElectrostatics() is required before build()"));
2356 newRunner.bonded_opt = bonded_opt_;
2360 GMX_THROW(gmx::APIError(
2361 "MdrunnerBuilder::addBondedTaskAssignment() is required before build()"));
2366 newRunner.update_opt = update_opt_;
2370 GMX_THROW(gmx::APIError(
2371 "MdrunnerBuilder::addUpdateTaskAssignment() is required before build() "));
2375 newRunner.restraintManager_ = std::make_unique<gmx::RestraintManager>();
2377 if (stopHandlerBuilder_)
2379 newRunner.stopHandlerBuilder_ = std::move(stopHandlerBuilder_);
2383 newRunner.stopHandlerBuilder_ = std::make_unique<StopHandlerBuilder>();
2389 void Mdrunner::BuilderImplementation::addHardwareDetectionResult(const gmx_hw_info_t* hwinfo)
2394 void Mdrunner::BuilderImplementation::addNonBonded(const char* nbpu_opt)
2396 nbpu_opt_ = nbpu_opt;
2399 void Mdrunner::BuilderImplementation::addPME(const char* pme_opt, const char* pme_fft_opt)
2402 pme_fft_opt_ = pme_fft_opt;
2405 void Mdrunner::BuilderImplementation::addBondedTaskAssignment(const char* bonded_opt)
2407 bonded_opt_ = bonded_opt;
2410 void Mdrunner::BuilderImplementation::addUpdateTaskAssignment(const char* update_opt)
2412 update_opt_ = update_opt;
2415 void Mdrunner::BuilderImplementation::addHardwareOptions(const gmx_hw_opt_t& hardwareOptions)
2417 hardwareOptions_ = hardwareOptions;
2420 void Mdrunner::BuilderImplementation::addFilenames(ArrayRef<const t_filenm> filenames)
2422 filenames_ = filenames;
2425 void Mdrunner::BuilderImplementation::addOutputEnvironment(gmx_output_env_t* outputEnvironment)
2427 outputEnvironment_ = outputEnvironment;
2430 void Mdrunner::BuilderImplementation::addLogFile(t_fileio* logFileHandle)
2432 logFileHandle_ = logFileHandle;
2435 void Mdrunner::BuilderImplementation::addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder)
2437 stopHandlerBuilder_ = std::move(builder);
2440 void Mdrunner::BuilderImplementation::addInput(SimulationInputHandle inputHolder)
2442 inputHolder_ = std::move(inputHolder);
2445 MdrunnerBuilder::MdrunnerBuilder(std::unique_ptr<MDModules> mdModules,
2446 compat::not_null<SimulationContext*> context) :
2447 impl_{ std::make_unique<Mdrunner::BuilderImplementation>(std::move(mdModules), context) }
2451 MdrunnerBuilder::~MdrunnerBuilder() = default;
2453 MdrunnerBuilder& MdrunnerBuilder::addHardwareDetectionResult(const gmx_hw_info_t* hwinfo)
2455 impl_->addHardwareDetectionResult(hwinfo);
2459 MdrunnerBuilder& MdrunnerBuilder::addSimulationMethod(const MdrunOptions& options,
2460 real forceWarningThreshold,
2461 const StartingBehavior startingBehavior)
2463 impl_->setExtraMdrunOptions(options, forceWarningThreshold, startingBehavior);
2467 MdrunnerBuilder& MdrunnerBuilder::addDomainDecomposition(const DomdecOptions& options)
2469 impl_->addDomdec(options);
2473 MdrunnerBuilder& MdrunnerBuilder::addNeighborList(int nstlist)
2475 impl_->addVerletList(nstlist);
2479 MdrunnerBuilder& MdrunnerBuilder::addReplicaExchange(const ReplicaExchangeParameters& params)
2481 impl_->addReplicaExchange(params);
2485 MdrunnerBuilder& MdrunnerBuilder::addNonBonded(const char* nbpu_opt)
2487 impl_->addNonBonded(nbpu_opt);
2491 MdrunnerBuilder& MdrunnerBuilder::addElectrostatics(const char* pme_opt, const char* pme_fft_opt)
2493 // The builder method may become more general in the future, but in this version,
2494 // parameters for PME electrostatics are both required and the only parameters
2496 if (pme_opt && pme_fft_opt)
2498 impl_->addPME(pme_opt, pme_fft_opt);
2503 gmx::InvalidInputError("addElectrostatics() arguments must be non-null pointers."));
2508 MdrunnerBuilder& MdrunnerBuilder::addBondedTaskAssignment(const char* bonded_opt)
2510 impl_->addBondedTaskAssignment(bonded_opt);
2514 MdrunnerBuilder& MdrunnerBuilder::addUpdateTaskAssignment(const char* update_opt)
2516 impl_->addUpdateTaskAssignment(update_opt);
2520 Mdrunner MdrunnerBuilder::build()
2522 return impl_->build();
2525 MdrunnerBuilder& MdrunnerBuilder::addHardwareOptions(const gmx_hw_opt_t& hardwareOptions)
2527 impl_->addHardwareOptions(hardwareOptions);
2531 MdrunnerBuilder& MdrunnerBuilder::addFilenames(ArrayRef<const t_filenm> filenames)
2533 impl_->addFilenames(filenames);
2537 MdrunnerBuilder& MdrunnerBuilder::addOutputEnvironment(gmx_output_env_t* outputEnvironment)
2539 impl_->addOutputEnvironment(outputEnvironment);
2543 MdrunnerBuilder& MdrunnerBuilder::addLogFile(t_fileio* logFileHandle)
2545 impl_->addLogFile(logFileHandle);
2549 MdrunnerBuilder& MdrunnerBuilder::addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder)
2551 impl_->addStopHandlerBuilder(std::move(builder));
2555 MdrunnerBuilder& MdrunnerBuilder::addInput(SimulationInputHandle input)
2557 impl_->addInput(std::move(input));
2561 MdrunnerBuilder::MdrunnerBuilder(MdrunnerBuilder&&) noexcept = default;
2563 MdrunnerBuilder& MdrunnerBuilder::operator=(MdrunnerBuilder&&) noexcept = default;