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, 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_gpu_program.h"
68 #include "gromacs/ewald/pme_only.h"
69 #include "gromacs/ewald/pme_pp_comm_gpu.h"
70 #include "gromacs/fileio/checkpoint.h"
71 #include "gromacs/fileio/gmxfio.h"
72 #include "gromacs/fileio/oenv.h"
73 #include "gromacs/fileio/tpxio.h"
74 #include "gromacs/gmxlib/network.h"
75 #include "gromacs/gmxlib/nrnb.h"
76 #include "gromacs/gpu_utils/device_context.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/printhardware.h"
82 #include "gromacs/imd/imd.h"
83 #include "gromacs/listed_forces/disre.h"
84 #include "gromacs/listed_forces/gpubonded.h"
85 #include "gromacs/listed_forces/listed_forces.h"
86 #include "gromacs/listed_forces/orires.h"
87 #include "gromacs/math/functions.h"
88 #include "gromacs/math/utilities.h"
89 #include "gromacs/math/vec.h"
90 #include "gromacs/mdlib/boxdeformation.h"
91 #include "gromacs/mdlib/broadcaststructs.h"
92 #include "gromacs/mdlib/calc_verletbuf.h"
93 #include "gromacs/mdlib/dispersioncorrection.h"
94 #include "gromacs/mdlib/enerdata_utils.h"
95 #include "gromacs/mdlib/force.h"
96 #include "gromacs/mdlib/forcerec.h"
97 #include "gromacs/mdlib/gmx_omp_nthreads.h"
98 #include "gromacs/mdlib/gpuforcereduction.h"
99 #include "gromacs/mdlib/makeconstraints.h"
100 #include "gromacs/mdlib/md_support.h"
101 #include "gromacs/mdlib/mdatoms.h"
102 #include "gromacs/mdlib/sighandler.h"
103 #include "gromacs/mdlib/stophandler.h"
104 #include "gromacs/mdlib/tgroup.h"
105 #include "gromacs/mdlib/updategroups.h"
106 #include "gromacs/mdlib/vsite.h"
107 #include "gromacs/mdrun/mdmodules.h"
108 #include "gromacs/mdrun/simulationcontext.h"
109 #include "gromacs/mdrun/simulationinput.h"
110 #include "gromacs/mdrun/simulationinputhandle.h"
111 #include "gromacs/mdrunutility/handlerestart.h"
112 #include "gromacs/mdrunutility/logging.h"
113 #include "gromacs/mdrunutility/multisim.h"
114 #include "gromacs/mdrunutility/printtime.h"
115 #include "gromacs/mdrunutility/threadaffinity.h"
116 #include "gromacs/mdtypes/checkpointdata.h"
117 #include "gromacs/mdtypes/commrec.h"
118 #include "gromacs/mdtypes/enerdata.h"
119 #include "gromacs/mdtypes/fcdata.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/observableshistory.h"
128 #include "gromacs/mdtypes/simulation_workload.h"
129 #include "gromacs/mdtypes/state.h"
130 #include "gromacs/mdtypes/state_propagator_data_gpu.h"
131 #include "gromacs/modularsimulator/modularsimulator.h"
132 #include "gromacs/nbnxm/gpu_data_mgmt.h"
133 #include "gromacs/nbnxm/nbnxm.h"
134 #include "gromacs/nbnxm/pairlist_tuning.h"
135 #include "gromacs/pbcutil/pbc.h"
136 #include "gromacs/pulling/output.h"
137 #include "gromacs/pulling/pull.h"
138 #include "gromacs/pulling/pull_rotation.h"
139 #include "gromacs/restraint/manager.h"
140 #include "gromacs/restraint/restraintmdmodule.h"
141 #include "gromacs/restraint/restraintpotential.h"
142 #include "gromacs/swap/swapcoords.h"
143 #include "gromacs/taskassignment/decidegpuusage.h"
144 #include "gromacs/taskassignment/decidesimulationworkload.h"
145 #include "gromacs/taskassignment/resourcedivision.h"
146 #include "gromacs/taskassignment/taskassignment.h"
147 #include "gromacs/taskassignment/usergpuids.h"
148 #include "gromacs/timing/gpu_timing.h"
149 #include "gromacs/timing/wallcycle.h"
150 #include "gromacs/timing/wallcyclereporting.h"
151 #include "gromacs/topology/mtop_util.h"
152 #include "gromacs/trajectory/trajectoryframe.h"
153 #include "gromacs/utility/basenetwork.h"
154 #include "gromacs/utility/cstringutil.h"
155 #include "gromacs/utility/exceptions.h"
156 #include "gromacs/utility/fatalerror.h"
157 #include "gromacs/utility/filestream.h"
158 #include "gromacs/utility/gmxassert.h"
159 #include "gromacs/utility/gmxmpi.h"
160 #include "gromacs/utility/keyvaluetree.h"
161 #include "gromacs/utility/logger.h"
162 #include "gromacs/utility/loggerbuilder.h"
163 #include "gromacs/utility/mdmodulenotification.h"
164 #include "gromacs/utility/physicalnodecommunicator.h"
165 #include "gromacs/utility/pleasecite.h"
166 #include "gromacs/utility/programcontext.h"
167 #include "gromacs/utility/smalloc.h"
168 #include "gromacs/utility/stringutil.h"
170 #include "isimulator.h"
171 #include "membedholder.h"
172 #include "replicaexchange.h"
173 #include "simulatorbuilder.h"
176 # include "corewrap.h"
183 /*! \brief Manage any development feature flag variables encountered
185 * The use of dev features indicated by environment variables is
186 * logged in order to ensure that runs with such features enabled can
187 * be identified from their log and standard output. Any cross
188 * dependencies are also checked, and if unsatisfied, a fatal error
191 * Note that some development features overrides are applied already here:
192 * the GPU communication flags are set to false in non-tMPI and non-CUDA builds.
194 * \param[in] mdlog Logger object.
195 * \param[in] useGpuForNonbonded True if the nonbonded task is offloaded in this run.
196 * \param[in] pmeRunMode The PME run mode for this run
197 * \returns The object populated with development feature flags.
199 static DevelopmentFeatureFlags manageDevelopmentFeatures(const gmx::MDLogger& mdlog,
200 const bool useGpuForNonbonded,
201 const PmeRunMode pmeRunMode)
203 DevelopmentFeatureFlags devFlags;
205 // Some builds of GCC 5 give false positive warnings that these
206 // getenv results are ignored when clearly they are used.
207 #pragma GCC diagnostic push
208 #pragma GCC diagnostic ignored "-Wunused-result"
210 devFlags.enableGpuBufferOps =
211 GMX_GPU_CUDA && useGpuForNonbonded && (getenv("GMX_USE_GPU_BUFFER_OPS") != nullptr);
212 devFlags.enableGpuHaloExchange = GMX_GPU_CUDA && GMX_THREAD_MPI && getenv("GMX_GPU_DD_COMMS") != nullptr;
213 devFlags.forceGpuUpdateDefault = (getenv("GMX_FORCE_UPDATE_DEFAULT_GPU") != nullptr);
214 devFlags.enableGpuPmePPComm =
215 GMX_GPU_CUDA && GMX_THREAD_MPI && getenv("GMX_GPU_PME_PP_COMMS") != nullptr;
217 #pragma GCC diagnostic pop
219 if (devFlags.enableGpuBufferOps)
221 GMX_LOG(mdlog.warning)
223 .appendTextFormatted(
224 "This run uses the 'GPU buffer ops' feature, enabled by the "
225 "GMX_USE_GPU_BUFFER_OPS environment variable.");
228 if (devFlags.forceGpuUpdateDefault)
230 GMX_LOG(mdlog.warning)
232 .appendTextFormatted(
233 "This run will default to '-update gpu' as requested by the "
234 "GMX_FORCE_UPDATE_DEFAULT_GPU environment variable. GPU update with domain "
235 "decomposition lacks substantial testing and should be used with caution.");
238 if (devFlags.enableGpuHaloExchange)
240 if (useGpuForNonbonded)
242 if (!devFlags.enableGpuBufferOps)
244 GMX_LOG(mdlog.warning)
246 .appendTextFormatted(
247 "Enabling GPU buffer operations required by GMX_GPU_DD_COMMS "
248 "(equivalent with GMX_USE_GPU_BUFFER_OPS=1).");
249 devFlags.enableGpuBufferOps = true;
251 GMX_LOG(mdlog.warning)
253 .appendTextFormatted(
254 "This run has requested the 'GPU halo exchange' feature, enabled by "
256 "GMX_GPU_DD_COMMS environment variable.");
260 GMX_LOG(mdlog.warning)
262 .appendTextFormatted(
263 "GMX_GPU_DD_COMMS environment variable detected, but the 'GPU "
264 "halo exchange' feature will not be enabled as nonbonded interactions "
265 "are not offloaded.");
266 devFlags.enableGpuHaloExchange = false;
270 if (devFlags.enableGpuPmePPComm)
272 if (pmeRunMode == PmeRunMode::GPU)
274 if (!devFlags.enableGpuBufferOps)
276 GMX_LOG(mdlog.warning)
278 .appendTextFormatted(
279 "Enabling GPU buffer operations required by GMX_GPU_PME_PP_COMMS "
280 "(equivalent with GMX_USE_GPU_BUFFER_OPS=1).");
281 devFlags.enableGpuBufferOps = true;
283 GMX_LOG(mdlog.warning)
285 .appendTextFormatted(
286 "This run uses the 'GPU PME-PP communications' feature, enabled "
287 "by the GMX_GPU_PME_PP_COMMS environment variable.");
291 std::string clarification;
292 if (pmeRunMode == PmeRunMode::Mixed)
295 "PME FFT and gather are not offloaded to the GPU (PME is running in mixed "
300 clarification = "PME is not offloaded to the GPU.";
302 GMX_LOG(mdlog.warning)
305 "GMX_GPU_PME_PP_COMMS environment variable detected, but the "
306 "'GPU PME-PP communications' feature was not enabled as "
308 devFlags.enableGpuPmePPComm = false;
315 /*! \brief Barrier for safe simultaneous thread access to mdrunner data
317 * Used to ensure that the master thread does not modify mdrunner during copy
318 * on the spawned threads. */
319 static void threadMpiMdrunnerAccessBarrier()
322 MPI_Barrier(MPI_COMM_WORLD);
326 Mdrunner Mdrunner::cloneOnSpawnedThread() const
328 auto newRunner = Mdrunner(std::make_unique<MDModules>());
330 // All runners in the same process share a restraint manager resource because it is
331 // part of the interface to the client code, which is associated only with the
332 // original thread. Handles to the same resources can be obtained by copy.
334 newRunner.restraintManager_ = std::make_unique<RestraintManager>(*restraintManager_);
337 // Copy members of master runner.
338 // \todo Replace with builder when Simulation context and/or runner phases are better defined.
339 // Ref https://gitlab.com/gromacs/gromacs/-/issues/2587 and https://gitlab.com/gromacs/gromacs/-/issues/2375
340 newRunner.hw_opt = hw_opt;
341 newRunner.filenames = filenames;
343 newRunner.oenv = oenv;
344 newRunner.mdrunOptions = mdrunOptions;
345 newRunner.domdecOptions = domdecOptions;
346 newRunner.nbpu_opt = nbpu_opt;
347 newRunner.pme_opt = pme_opt;
348 newRunner.pme_fft_opt = pme_fft_opt;
349 newRunner.bonded_opt = bonded_opt;
350 newRunner.update_opt = update_opt;
351 newRunner.nstlist_cmdline = nstlist_cmdline;
352 newRunner.replExParams = replExParams;
353 newRunner.pforce = pforce;
354 // Give the spawned thread the newly created valid communicator
355 // for the simulation.
356 newRunner.libraryWorldCommunicator = MPI_COMM_WORLD;
357 newRunner.simulationCommunicator = MPI_COMM_WORLD;
359 newRunner.startingBehavior = startingBehavior;
360 newRunner.stopHandlerBuilder_ = std::make_unique<StopHandlerBuilder>(*stopHandlerBuilder_);
361 newRunner.inputHolder_ = inputHolder_;
363 threadMpiMdrunnerAccessBarrier();
368 /*! \brief The callback used for running on spawned threads.
370 * Obtains the pointer to the master mdrunner object from the one
371 * argument permitted to the thread-launch API call, copies it to make
372 * a new runner for this thread, reinitializes necessary data, and
373 * proceeds to the simulation. */
374 static void mdrunner_start_fn(const void* arg)
378 auto masterMdrunner = reinterpret_cast<const gmx::Mdrunner*>(arg);
379 /* copy the arg list to make sure that it's thread-local. This
380 doesn't copy pointed-to items, of course; fnm, cr and fplog
381 are reset in the call below, all others should be const. */
382 gmx::Mdrunner mdrunner = masterMdrunner->cloneOnSpawnedThread();
385 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
389 void Mdrunner::spawnThreads(int numThreadsToLaunch)
392 /* now spawn new threads that start mdrunner_start_fn(), while
393 the main thread returns. Thread affinity is handled later. */
394 if (tMPI_Init_fn(TRUE, numThreadsToLaunch, TMPI_AFFINITY_NONE, mdrunner_start_fn,
395 static_cast<const void*>(this))
398 GMX_THROW(gmx::InternalError("Failed to spawn thread-MPI threads"));
401 // Give the master thread the newly created valid communicator for
403 libraryWorldCommunicator = MPI_COMM_WORLD;
404 simulationCommunicator = MPI_COMM_WORLD;
405 threadMpiMdrunnerAccessBarrier();
407 GMX_UNUSED_VALUE(numThreadsToLaunch);
408 GMX_UNUSED_VALUE(mdrunner_start_fn);
414 /*! \brief Initialize variables for Verlet scheme simulation */
415 static void prepare_verlet_scheme(FILE* fplog,
419 const gmx_mtop_t* mtop,
421 bool makeGpuPairList,
422 const gmx::CpuInfo& cpuinfo)
424 // We checked the cut-offs in grompp, but double-check here.
425 // We have PME+LJcutoff kernels for rcoulomb>rvdw.
426 if (EEL_PME_EWALD(ir->coulombtype) && ir->vdwtype == eelCUT)
428 GMX_RELEASE_ASSERT(ir->rcoulomb >= ir->rvdw,
429 "With Verlet lists and PME we should have rcoulomb>=rvdw");
433 GMX_RELEASE_ASSERT(ir->rcoulomb == ir->rvdw,
434 "With Verlet lists and no PME rcoulomb and rvdw should be identical");
436 /* For NVE simulations, we will retain the initial list buffer */
437 if (EI_DYNAMICS(ir->eI) && ir->verletbuf_tol > 0 && !(EI_MD(ir->eI) && ir->etc == etcNO))
439 /* Update the Verlet buffer size for the current run setup */
441 /* Here we assume SIMD-enabled kernels are being used. But as currently
442 * calc_verlet_buffer_size gives the same results for 4x8 and 4x4
443 * and 4x2 gives a larger buffer than 4x4, this is ok.
445 ListSetupType listType =
446 (makeGpuPairList ? ListSetupType::Gpu : ListSetupType::CpuSimdWhenSupported);
447 VerletbufListSetup listSetup = verletbufGetSafeListSetup(listType);
449 const real rlist_new =
450 calcVerletBufferSize(*mtop, det(box), *ir, ir->nstlist, ir->nstlist - 1, -1, listSetup);
452 if (rlist_new != ir->rlist)
454 if (fplog != nullptr)
457 "\nChanging rlist from %g to %g for non-bonded %dx%d atom kernels\n\n",
458 ir->rlist, rlist_new, listSetup.cluster_size_i, listSetup.cluster_size_j);
460 ir->rlist = rlist_new;
464 if (nstlist_cmdline > 0 && (!EI_DYNAMICS(ir->eI) || ir->verletbuf_tol <= 0))
466 gmx_fatal(FARGS, "Can not set nstlist without %s",
467 !EI_DYNAMICS(ir->eI) ? "dynamics" : "verlet-buffer-tolerance");
470 if (EI_DYNAMICS(ir->eI))
472 /* Set or try nstlist values */
473 increaseNstlist(fplog, cr, ir, nstlist_cmdline, mtop, box, makeGpuPairList, cpuinfo);
477 /*! \brief Override the nslist value in inputrec
479 * with value passed on the command line (if any)
481 static void override_nsteps_cmdline(const gmx::MDLogger& mdlog, int64_t nsteps_cmdline, t_inputrec* ir)
485 /* override with anything else than the default -2 */
486 if (nsteps_cmdline > -2)
488 char sbuf_steps[STEPSTRSIZE];
489 char sbuf_msg[STRLEN];
491 ir->nsteps = nsteps_cmdline;
492 if (EI_DYNAMICS(ir->eI) && nsteps_cmdline != -1)
495 "Overriding nsteps with value passed on the command line: %s steps, %.3g ps",
496 gmx_step_str(nsteps_cmdline, sbuf_steps), fabs(nsteps_cmdline * ir->delta_t));
500 sprintf(sbuf_msg, "Overriding nsteps with value passed on the command line: %s steps",
501 gmx_step_str(nsteps_cmdline, sbuf_steps));
504 GMX_LOG(mdlog.warning).asParagraph().appendText(sbuf_msg);
506 else if (nsteps_cmdline < -2)
508 gmx_fatal(FARGS, "Invalid nsteps value passed on the command line: %" PRId64, nsteps_cmdline);
510 /* Do nothing if nsteps_cmdline == -2 */
516 /*! \brief Return whether GPU acceleration of nonbondeds is supported with the given settings.
518 * If not, and if a warning may be issued, logs a warning about
519 * falling back to CPU code. With thread-MPI, only the first
520 * call to this function should have \c issueWarning true. */
521 static bool gpuAccelerationOfNonbondedIsUseful(const MDLogger& mdlog, const t_inputrec& ir, bool issueWarning)
523 bool gpuIsUseful = true;
526 if (ir.opts.ngener - ir.nwall > 1)
528 /* The GPU code does not support more than one energy group.
529 * If the user requested GPUs explicitly, a fatal error is given later.
533 "Multiple energy groups is not implemented for GPUs, falling back to the CPU. "
534 "For better performance, run on the GPU without energy groups and then do "
535 "gmx mdrun -rerun option on the trajectory with an energy group .tpr file.";
541 warning = "TPI is not implemented for GPUs.";
544 if (!gpuIsUseful && issueWarning)
546 GMX_LOG(mdlog.warning).asParagraph().appendText(warning);
552 //! Initializes the logger for mdrun.
553 static gmx::LoggerOwner buildLogger(FILE* fplog, const bool isSimulationMasterRank)
555 gmx::LoggerBuilder builder;
556 if (fplog != nullptr)
558 builder.addTargetFile(gmx::MDLogger::LogLevel::Info, fplog);
560 if (isSimulationMasterRank)
562 builder.addTargetStream(gmx::MDLogger::LogLevel::Warning, &gmx::TextOutputFile::standardError());
564 return builder.build();
567 //! Make a TaskTarget from an mdrun argument string.
568 static TaskTarget findTaskTarget(const char* optionString)
570 TaskTarget returnValue = TaskTarget::Auto;
572 if (strncmp(optionString, "auto", 3) == 0)
574 returnValue = TaskTarget::Auto;
576 else if (strncmp(optionString, "cpu", 3) == 0)
578 returnValue = TaskTarget::Cpu;
580 else if (strncmp(optionString, "gpu", 3) == 0)
582 returnValue = TaskTarget::Gpu;
586 GMX_ASSERT(false, "Option string should have been checked for sanity already");
592 //! Finish run, aggregate data to print performance info.
593 static void finish_run(FILE* fplog,
594 const gmx::MDLogger& mdlog,
596 const t_inputrec* inputrec,
598 gmx_wallcycle_t wcycle,
599 gmx_walltime_accounting_t walltime_accounting,
600 nonbonded_verlet_t* nbv,
601 const gmx_pme_t* pme,
605 double nbfs = 0, mflop = 0;
606 double elapsed_time, elapsed_time_over_all_ranks, elapsed_time_over_all_threads,
607 elapsed_time_over_all_threads_over_all_ranks;
608 /* Control whether it is valid to print a report. Only the
609 simulation master may print, but it should not do so if the run
610 terminated e.g. before a scheduled reset step. This is
611 complicated by the fact that PME ranks are unaware of the
612 reason why they were sent a pmerecvqxFINISH. To avoid
613 communication deadlocks, we always do the communication for the
614 report, even if we've decided not to write the report, because
615 how long it takes to finish the run is not important when we've
616 decided not to report on the simulation performance.
618 Further, we only report performance for dynamical integrators,
619 because those are the only ones for which we plan to
620 consider doing any optimizations. */
621 bool printReport = EI_DYNAMICS(inputrec->eI) && SIMMASTER(cr);
623 if (printReport && !walltime_accounting_get_valid_finish(walltime_accounting))
625 GMX_LOG(mdlog.warning)
627 .appendText("Simulation ended prematurely, no performance report will be written.");
632 std::unique_ptr<t_nrnb> nrnbTotalStorage;
635 nrnbTotalStorage = std::make_unique<t_nrnb>();
636 nrnb_tot = nrnbTotalStorage.get();
638 MPI_Allreduce(nrnb->n, nrnb_tot->n, eNRNB, MPI_DOUBLE, MPI_SUM, cr->mpi_comm_mysim);
646 elapsed_time = walltime_accounting_get_time_since_reset(walltime_accounting);
647 elapsed_time_over_all_threads =
648 walltime_accounting_get_time_since_reset_over_all_threads(walltime_accounting);
652 /* reduce elapsed_time over all MPI ranks in the current simulation */
653 MPI_Allreduce(&elapsed_time, &elapsed_time_over_all_ranks, 1, MPI_DOUBLE, MPI_SUM,
655 elapsed_time_over_all_ranks /= cr->nnodes;
656 /* Reduce elapsed_time_over_all_threads over all MPI ranks in the
657 * current simulation. */
658 MPI_Allreduce(&elapsed_time_over_all_threads, &elapsed_time_over_all_threads_over_all_ranks,
659 1, MPI_DOUBLE, MPI_SUM, cr->mpi_comm_mysim);
664 elapsed_time_over_all_ranks = elapsed_time;
665 elapsed_time_over_all_threads_over_all_ranks = elapsed_time_over_all_threads;
670 print_flop(fplog, nrnb_tot, &nbfs, &mflop);
673 if (thisRankHasDuty(cr, DUTY_PP) && DOMAINDECOMP(cr))
675 print_dd_statistics(cr, inputrec, fplog);
678 /* TODO Move the responsibility for any scaling by thread counts
679 * to the code that handled the thread region, so that there's a
680 * mechanism to keep cycle counting working during the transition
681 * to task parallelism. */
682 int nthreads_pp = gmx_omp_nthreads_get(emntNonbonded);
683 int nthreads_pme = gmx_omp_nthreads_get(emntPME);
684 wallcycle_scale_by_num_threads(wcycle, thisRankHasDuty(cr, DUTY_PME) && !thisRankHasDuty(cr, DUTY_PP),
685 nthreads_pp, nthreads_pme);
686 auto cycle_sum(wallcycle_sum(cr, wcycle));
690 auto nbnxn_gpu_timings =
691 (nbv != nullptr && nbv->useGpu()) ? Nbnxm::gpu_get_timings(nbv->gpu_nbv) : nullptr;
692 gmx_wallclock_gpu_pme_t pme_gpu_timings = {};
694 if (pme_gpu_task_enabled(pme))
696 pme_gpu_get_timings(pme, &pme_gpu_timings);
698 wallcycle_print(fplog, mdlog, cr->nnodes, cr->npmenodes, nthreads_pp, nthreads_pme,
699 elapsed_time_over_all_ranks, wcycle, cycle_sum, nbnxn_gpu_timings,
702 if (EI_DYNAMICS(inputrec->eI))
704 delta_t = inputrec->delta_t;
709 print_perf(fplog, elapsed_time_over_all_threads_over_all_ranks, elapsed_time_over_all_ranks,
710 walltime_accounting_get_nsteps_done_since_reset(walltime_accounting),
711 delta_t, nbfs, mflop);
715 print_perf(stderr, elapsed_time_over_all_threads_over_all_ranks, elapsed_time_over_all_ranks,
716 walltime_accounting_get_nsteps_done_since_reset(walltime_accounting),
717 delta_t, nbfs, mflop);
722 int Mdrunner::mdrunner()
725 t_forcerec* fr = nullptr;
726 real ewaldcoeff_q = 0;
727 real ewaldcoeff_lj = 0;
728 int nChargePerturbed = -1, nTypePerturbed = 0;
729 gmx_wallcycle_t wcycle;
730 gmx_walltime_accounting_t walltime_accounting = nullptr;
731 MembedHolder membedHolder(filenames.size(), filenames.data());
732 gmx_hw_info_t* hwinfo = nullptr;
734 /* CAUTION: threads may be started later on in this function, so
735 cr doesn't reflect the final parallel state right now */
738 /* TODO: inputrec should tell us whether we use an algorithm, not a file option */
739 const bool doEssentialDynamics = opt2bSet("-ei", filenames.size(), filenames.data());
740 const bool doRerun = mdrunOptions.rerun;
742 // Handle task-assignment related user options.
743 EmulateGpuNonbonded emulateGpuNonbonded =
744 (getenv("GMX_EMULATE_GPU") != nullptr ? EmulateGpuNonbonded::Yes : EmulateGpuNonbonded::No);
746 std::vector<int> userGpuTaskAssignment;
749 userGpuTaskAssignment = parseUserTaskAssignmentString(hw_opt.userGpuTaskAssignment);
751 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
752 auto nonbondedTarget = findTaskTarget(nbpu_opt);
753 auto pmeTarget = findTaskTarget(pme_opt);
754 auto pmeFftTarget = findTaskTarget(pme_fft_opt);
755 auto bondedTarget = findTaskTarget(bonded_opt);
756 auto updateTarget = findTaskTarget(update_opt);
758 FILE* fplog = nullptr;
759 // If we are appending, we don't write log output because we need
760 // to check that the old log file matches what the checkpoint file
761 // expects. Otherwise, we should start to write log output now if
762 // there is a file ready for it.
763 if (logFileHandle != nullptr && startingBehavior != StartingBehavior::RestartWithAppending)
765 fplog = gmx_fio_getfp(logFileHandle);
767 const bool isSimulationMasterRank = findIsSimulationMasterRank(ms, simulationCommunicator);
768 gmx::LoggerOwner logOwner(buildLogger(fplog, isSimulationMasterRank));
769 gmx::MDLogger mdlog(logOwner.logger());
771 // TODO The thread-MPI master rank makes a working
772 // PhysicalNodeCommunicator here, but it gets rebuilt by all ranks
773 // after the threads have been launched. This works because no use
774 // is made of that communicator until after the execution paths
775 // have rejoined. But it is likely that we can improve the way
776 // this is expressed, e.g. by expressly running detection only the
777 // master rank for thread-MPI, rather than relying on the mutex
778 // and reference count.
779 PhysicalNodeCommunicator physicalNodeComm(libraryWorldCommunicator, gmx_physicalnode_id_hash());
780 hwinfo = gmx_detect_hardware(mdlog, physicalNodeComm);
782 gmx_print_detected_hardware(fplog, isSimulationMasterRank && isMasterSim(ms), mdlog, hwinfo);
784 std::vector<int> gpuIdsToUse = makeGpuIdsToUse(hwinfo->deviceInfoList, hw_opt.gpuIdsAvailable);
786 // Print citation requests after all software/hardware printing
787 pleaseCiteGromacs(fplog);
789 // Note: legacy program logic relies on checking whether these pointers are assigned.
790 // Objects may or may not be allocated later.
791 std::unique_ptr<t_inputrec> inputrec;
792 std::unique_ptr<t_state> globalState;
794 auto partialDeserializedTpr = std::make_unique<PartialDeserializedTprFile>();
796 if (isSimulationMasterRank)
798 // Allocate objects to be initialized by later function calls.
799 /* Only the master rank has the global state */
800 globalState = std::make_unique<t_state>();
801 inputrec = std::make_unique<t_inputrec>();
803 /* Read (nearly) all data required for the simulation
804 * and keep the partly serialized tpr contents to send to other ranks later
806 applyGlobalSimulationState(*inputHolder_.get(), partialDeserializedTpr.get(),
807 globalState.get(), inputrec.get(), &mtop);
810 /* Check and update the hardware options for internal consistency */
811 checkAndUpdateHardwareOptions(mdlog, &hw_opt, isSimulationMasterRank, domdecOptions.numPmeRanks,
814 if (GMX_THREAD_MPI && isSimulationMasterRank)
816 bool useGpuForNonbonded = false;
817 bool useGpuForPme = false;
820 GMX_RELEASE_ASSERT(inputrec != nullptr, "Keep the compiler happy");
822 // If the user specified the number of ranks, then we must
823 // respect that, but in default mode, we need to allow for
824 // the number of GPUs to choose the number of ranks.
825 auto canUseGpuForNonbonded = buildSupportsNonbondedOnGpu(nullptr);
826 useGpuForNonbonded = decideWhetherToUseGpusForNonbondedWithThreadMpi(
827 nonbondedTarget, gpuIdsToUse, userGpuTaskAssignment, emulateGpuNonbonded,
828 canUseGpuForNonbonded,
829 gpuAccelerationOfNonbondedIsUseful(mdlog, *inputrec, GMX_THREAD_MPI),
830 hw_opt.nthreads_tmpi);
831 useGpuForPme = decideWhetherToUseGpusForPmeWithThreadMpi(
832 useGpuForNonbonded, pmeTarget, gpuIdsToUse, userGpuTaskAssignment, *hwinfo,
833 *inputrec, hw_opt.nthreads_tmpi, domdecOptions.numPmeRanks);
835 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
837 /* Determine how many thread-MPI ranks to start.
839 * TODO Over-writing the user-supplied value here does
840 * prevent any possible subsequent checks from working
842 hw_opt.nthreads_tmpi =
843 get_nthreads_mpi(hwinfo, &hw_opt, gpuIdsToUse, useGpuForNonbonded, useGpuForPme,
844 inputrec.get(), &mtop, mdlog, membedHolder.doMembed());
846 // Now start the threads for thread MPI.
847 spawnThreads(hw_opt.nthreads_tmpi);
848 // The spawned threads enter mdrunner() and execution of
849 // master and spawned threads joins at the end of this block.
851 PhysicalNodeCommunicator(libraryWorldCommunicator, gmx_physicalnode_id_hash());
854 GMX_RELEASE_ASSERT(ms || simulationCommunicator != MPI_COMM_NULL,
855 "Must have valid communicator unless running a multi-simulation");
856 CommrecHandle crHandle = init_commrec(simulationCommunicator);
857 t_commrec* cr = crHandle.get();
858 GMX_RELEASE_ASSERT(cr != nullptr, "Must have valid commrec");
862 /* now broadcast everything to the non-master nodes/threads: */
863 if (!isSimulationMasterRank)
865 // Until now, only the master rank has a non-null pointer.
866 // On non-master ranks, allocate the object that will receive data in the following call.
867 inputrec = std::make_unique<t_inputrec>();
869 init_parallel(cr->mpiDefaultCommunicator, MASTER(cr), inputrec.get(), &mtop,
870 partialDeserializedTpr.get());
872 GMX_RELEASE_ASSERT(inputrec != nullptr, "All ranks should have a valid inputrec now");
873 partialDeserializedTpr.reset(nullptr);
875 // Now the number of ranks is known to all ranks, and each knows
876 // the inputrec read by the master rank. The ranks can now all run
877 // the task-deciding functions and will agree on the result
878 // without needing to communicate.
879 const bool useDomainDecomposition = (PAR(cr) && !(EI_TPI(inputrec->eI) || inputrec->eI == eiNM));
881 // Note that these variables describe only their own node.
883 // Note that when bonded interactions run on a GPU they always run
884 // alongside a nonbonded task, so do not influence task assignment
885 // even though they affect the force calculation workload.
886 bool useGpuForNonbonded = false;
887 bool useGpuForPme = false;
888 bool useGpuForBonded = false;
889 bool useGpuForUpdate = false;
890 bool gpusWereDetected = hwinfo->ngpu_compatible_tot > 0;
893 // It's possible that there are different numbers of GPUs on
894 // different nodes, which is the user's responsibility to
895 // handle. If unsuitable, we will notice that during task
897 auto canUseGpuForNonbonded = buildSupportsNonbondedOnGpu(nullptr);
898 useGpuForNonbonded = decideWhetherToUseGpusForNonbonded(
899 nonbondedTarget, userGpuTaskAssignment, emulateGpuNonbonded, canUseGpuForNonbonded,
900 gpuAccelerationOfNonbondedIsUseful(mdlog, *inputrec, !GMX_THREAD_MPI), gpusWereDetected);
901 useGpuForPme = decideWhetherToUseGpusForPme(
902 useGpuForNonbonded, pmeTarget, userGpuTaskAssignment, *hwinfo, *inputrec,
903 cr->sizeOfDefaultCommunicator, domdecOptions.numPmeRanks, gpusWereDetected);
904 auto canUseGpuForBonded = buildSupportsGpuBondeds(nullptr)
905 && inputSupportsGpuBondeds(*inputrec, mtop, nullptr);
906 useGpuForBonded = decideWhetherToUseGpusForBonded(
907 useGpuForNonbonded, useGpuForPme, bondedTarget, canUseGpuForBonded,
908 EVDW_PME(inputrec->vdwtype), EEL_PME_EWALD(inputrec->coulombtype),
909 domdecOptions.numPmeRanks, gpusWereDetected);
911 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
913 const PmeRunMode pmeRunMode = determinePmeRunMode(useGpuForPme, pmeFftTarget, *inputrec);
915 // Initialize development feature flags that enabled by environment variable
916 // and report those features that are enabled.
917 const DevelopmentFeatureFlags devFlags =
918 manageDevelopmentFeatures(mdlog, useGpuForNonbonded, pmeRunMode);
920 const bool useModularSimulator =
921 checkUseModularSimulator(false, inputrec.get(), doRerun, mtop, ms, replExParams,
922 nullptr, doEssentialDynamics, membedHolder.doMembed());
925 // TODO: hide restraint implementation details from Mdrunner.
926 // There is nothing unique about restraints at this point as far as the
927 // Mdrunner is concerned. The Mdrunner should just be getting a sequence of
928 // factory functions from the SimulationContext on which to call mdModules_->add().
929 // TODO: capture all restraints into a single RestraintModule, passed to the runner builder.
930 for (auto&& restraint : restraintManager_->getRestraints())
932 auto module = RestraintMDModule::create(restraint, restraint->sites());
933 mdModules_->add(std::move(module));
936 // TODO: Error handling
937 mdModules_->assignOptionsToModules(*inputrec->params, nullptr);
938 // now that the MdModules know their options, they know which callbacks to sign up to
939 mdModules_->subscribeToSimulationSetupNotifications();
940 const auto& mdModulesNotifier = mdModules_->notifier().simulationSetupNotifications_;
942 if (inputrec->internalParameters != nullptr)
944 mdModulesNotifier.notify(*inputrec->internalParameters);
947 if (fplog != nullptr)
949 pr_inputrec(fplog, 0, "Input Parameters", inputrec.get(), FALSE);
950 fprintf(fplog, "\n");
955 /* In rerun, set velocities to zero if present */
956 if (doRerun && ((globalState->flags & (1 << estV)) != 0))
958 // rerun does not use velocities
962 "Rerun trajectory contains velocities. Rerun does only evaluate "
963 "potential energy and forces. The velocities will be ignored.");
964 for (int i = 0; i < globalState->natoms; i++)
966 clear_rvec(globalState->v[i]);
968 globalState->flags &= ~(1 << estV);
971 /* now make sure the state is initialized and propagated */
972 set_state_entries(globalState.get(), inputrec.get(), useModularSimulator);
975 /* NM and TPI parallelize over force/energy calculations, not atoms,
976 * so we need to initialize and broadcast the global state.
978 if (inputrec->eI == eiNM || inputrec->eI == eiTPI)
982 globalState = std::make_unique<t_state>();
984 broadcastStateWithoutDynamics(cr->mpiDefaultCommunicator, DOMAINDECOMP(cr), PAR(cr),
988 /* A parallel command line option consistency check that we can
989 only do after any threads have started. */
991 && (domdecOptions.numCells[XX] > 1 || domdecOptions.numCells[YY] > 1
992 || domdecOptions.numCells[ZZ] > 1 || domdecOptions.numPmeRanks > 0))
995 "The -dd or -npme option request a parallel simulation, "
997 "but %s was compiled without threads or MPI enabled",
998 output_env_get_program_display_name(oenv));
1000 "but the number of MPI-threads (option -ntmpi) is not set or is 1");
1002 "but %s was not started through mpirun/mpiexec or only one rank was requested "
1003 "through mpirun/mpiexec",
1004 output_env_get_program_display_name(oenv));
1008 if (doRerun && (EI_ENERGY_MINIMIZATION(inputrec->eI) || eiNM == inputrec->eI))
1011 "The .mdp file specified an energy mininization or normal mode algorithm, and "
1012 "these are not compatible with mdrun -rerun");
1015 if (!(EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype)))
1017 if (domdecOptions.numPmeRanks > 0)
1019 gmx_fatal_collective(FARGS, cr->mpiDefaultCommunicator, MASTER(cr),
1020 "PME-only ranks are requested, but the system does not use PME "
1021 "for electrostatics or LJ");
1024 domdecOptions.numPmeRanks = 0;
1027 if (useGpuForNonbonded && domdecOptions.numPmeRanks < 0)
1029 /* With NB GPUs we don't automatically use PME-only CPU ranks. PME ranks can
1030 * improve performance with many threads per GPU, since our OpenMP
1031 * scaling is bad, but it's difficult to automate the setup.
1033 domdecOptions.numPmeRanks = 0;
1037 if (domdecOptions.numPmeRanks < 0)
1039 domdecOptions.numPmeRanks = 0;
1040 // TODO possibly print a note that one can opt-in for a separate PME GPU rank?
1044 GMX_RELEASE_ASSERT(domdecOptions.numPmeRanks <= 1,
1045 "PME GPU decomposition is not supported");
1052 fcRegisterSteps(inputrec->nsteps, inputrec->init_step);
1056 /* NMR restraints must be initialized before load_checkpoint,
1057 * since with time averaging the history is added to t_state.
1058 * For proper consistency check we therefore need to extend
1060 * So the PME-only nodes (if present) will also initialize
1061 * the distance restraints.
1064 /* This needs to be called before read_checkpoint to extend the state */
1065 t_disresdata* disresdata;
1066 snew(disresdata, 1);
1067 init_disres(fplog, &mtop, inputrec.get(), DisResRunMode::MDRun,
1068 MASTER(cr) ? DDRole::Master : DDRole::Agent,
1069 PAR(cr) ? NumRanks::Multiple : NumRanks::Single, cr->mpi_comm_mysim, ms, disresdata,
1070 globalState.get(), replExParams.exchangeInterval > 0);
1072 t_oriresdata* oriresdata;
1073 snew(oriresdata, 1);
1074 init_orires(fplog, &mtop, inputrec.get(), cr, ms, globalState.get(), oriresdata);
1076 auto deform = prepareBoxDeformation(
1077 globalState != nullptr ? globalState->box : box, MASTER(cr) ? DDRole::Master : DDRole::Agent,
1078 PAR(cr) ? NumRanks::Multiple : NumRanks::Single, cr->mpi_comm_mygroup, *inputrec);
1080 ObservablesHistory observablesHistory = {};
1082 auto modularSimulatorCheckpointData = std::make_unique<ReadCheckpointDataHolder>();
1083 if (startingBehavior != StartingBehavior::NewSimulation)
1085 /* Check if checkpoint file exists before doing continuation.
1086 * This way we can use identical input options for the first and subsequent runs...
1088 if (mdrunOptions.numStepsCommandline > -2)
1090 /* Temporarily set the number of steps to unlimited to avoid
1091 * triggering the nsteps check in load_checkpoint().
1092 * This hack will go away soon when the -nsteps option is removed.
1094 inputrec->nsteps = -1;
1097 // Finish applying initial simulation state information from external sources on all ranks.
1098 // Reconcile checkpoint file data with Mdrunner state established up to this point.
1099 applyLocalState(*inputHolder_.get(), logFileHandle, cr, domdecOptions.numCells,
1100 inputrec.get(), globalState.get(), &observablesHistory,
1101 mdrunOptions.reproducible, mdModules_->notifier(),
1102 modularSimulatorCheckpointData.get(), useModularSimulator);
1103 // TODO: (#3652) Synchronize filesystem state, SimulationInput contents, and program
1105 // on all code paths.
1106 // Write checkpoint or provide hook to update SimulationInput.
1107 // If there was a checkpoint file, SimulationInput contains more information
1108 // than if there wasn't. At this point, we have synchronized the in-memory
1109 // state with the filesystem state only for restarted simulations. We should
1110 // be calling applyLocalState unconditionally and expect that the completeness
1111 // of SimulationInput is not dependent on its creation method.
1113 if (startingBehavior == StartingBehavior::RestartWithAppending && logFileHandle)
1115 // Now we can start normal logging to the truncated log file.
1116 fplog = gmx_fio_getfp(logFileHandle);
1117 prepareLogAppending(fplog);
1118 logOwner = buildLogger(fplog, MASTER(cr));
1119 mdlog = logOwner.logger();
1123 if (mdrunOptions.numStepsCommandline > -2)
1128 "The -nsteps functionality is deprecated, and may be removed in a future "
1130 "Consider using gmx convert-tpr -nsteps or changing the appropriate .mdp "
1133 /* override nsteps with value set on the commandline */
1134 override_nsteps_cmdline(mdlog, mdrunOptions.numStepsCommandline, inputrec.get());
1136 if (isSimulationMasterRank)
1138 copy_mat(globalState->box, box);
1143 gmx_bcast(sizeof(box), box, cr->mpiDefaultCommunicator);
1146 if (inputrec->cutoff_scheme != ecutsVERLET)
1149 "This group-scheme .tpr file can no longer be run by mdrun. Please update to the "
1150 "Verlet scheme, or use an earlier version of GROMACS if necessary.");
1152 /* Update rlist and nstlist. */
1153 /* Note: prepare_verlet_scheme is calling increaseNstlist(...), which (while attempting to
1154 * increase rlist) tries to check if the newly chosen value fits with the DD scheme. As this is
1155 * run before any DD scheme is set up, this check is never executed. See #3334 for more details.
1157 prepare_verlet_scheme(fplog, cr, inputrec.get(), nstlist_cmdline, &mtop, box,
1158 useGpuForNonbonded || (emulateGpuNonbonded == EmulateGpuNonbonded::Yes),
1161 // This builder is necessary while we have multi-part construction
1162 // of DD. Before DD is constructed, we use the existence of
1163 // the builder object to indicate that further construction of DD
1165 std::unique_ptr<DomainDecompositionBuilder> ddBuilder;
1166 if (useDomainDecomposition)
1168 ddBuilder = std::make_unique<DomainDecompositionBuilder>(
1169 mdlog, cr, domdecOptions, mdrunOptions, mtop, *inputrec, box,
1170 positionsFromStatePointer(globalState.get()));
1174 /* PME, if used, is done on all nodes with 1D decomposition */
1175 cr->nnodes = cr->sizeOfDefaultCommunicator;
1176 cr->sim_nodeid = cr->rankInDefaultCommunicator;
1177 cr->nodeid = cr->rankInDefaultCommunicator;
1179 cr->duty = (DUTY_PP | DUTY_PME);
1181 if (inputrec->pbcType == PbcType::Screw)
1183 gmx_fatal(FARGS, "pbc=screw is only implemented with domain decomposition");
1187 // Produce the task assignment for this rank - done after DD is constructed
1188 GpuTaskAssignments gpuTaskAssignments = GpuTaskAssignmentsBuilder::build(
1189 gpuIdsToUse, userGpuTaskAssignment, *hwinfo, simulationCommunicator, physicalNodeComm,
1190 nonbondedTarget, pmeTarget, bondedTarget, updateTarget, useGpuForNonbonded,
1191 useGpuForPme, thisRankHasDuty(cr, DUTY_PP),
1192 // TODO cr->duty & DUTY_PME should imply that a PME
1193 // algorithm is active, but currently does not.
1194 EEL_PME(inputrec->coulombtype) && thisRankHasDuty(cr, DUTY_PME));
1196 // Get the device handles for the modules, nullptr when no task is assigned.
1198 DeviceInformation* deviceInfo = gpuTaskAssignments.initDevice(&deviceId);
1200 // timing enabling - TODO put this in gpu_utils (even though generally this is just option handling?)
1201 bool useTiming = true;
1205 /* WARNING: CUDA timings are incorrect with multiple streams.
1206 * This is the main reason why they are disabled by default.
1208 // TODO: Consider turning on by default when we can detect nr of streams.
1209 useTiming = (getenv("GMX_ENABLE_GPU_TIMING") != nullptr);
1211 else if (GMX_GPU_OPENCL)
1213 useTiming = (getenv("GMX_DISABLE_GPU_TIMING") == nullptr);
1216 // TODO Currently this is always built, yet DD partition code
1217 // checks if it is built before using it. Probably it should
1218 // become an MDModule that is made only when another module
1219 // requires it (e.g. pull, CompEl, density fitting), so that we
1220 // don't update the local atom sets unilaterally every step.
1221 LocalAtomSetManager atomSets;
1224 // TODO Pass the GPU streams to ddBuilder to use in buffer
1225 // transfers (e.g. halo exchange)
1226 cr->dd = ddBuilder->build(&atomSets);
1227 // The builder's job is done, so destruct it
1228 ddBuilder.reset(nullptr);
1229 // Note that local state still does not exist yet.
1232 // The GPU update is decided here because we need to know whether the constraints or
1233 // SETTLEs can span accross the domain borders (i.e. whether or not update groups are
1234 // defined). This is only known after DD is initialized, hence decision on using GPU
1235 // update is done so late.
1238 const bool useUpdateGroups = cr->dd ? ddUsesUpdateGroups(*cr->dd) : false;
1240 useGpuForUpdate = decideWhetherToUseGpuForUpdate(
1241 useDomainDecomposition, useUpdateGroups, pmeRunMode, domdecOptions.numPmeRanks > 0,
1242 useGpuForNonbonded, updateTarget, gpusWereDetected, *inputrec, mtop,
1243 doEssentialDynamics, gmx_mtop_ftype_count(mtop, F_ORIRES) > 0,
1244 replExParams.exchangeInterval > 0, doRerun, devFlags, mdlog);
1246 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1248 const bool printHostName = (cr->nnodes > 1);
1249 gpuTaskAssignments.reportGpuUsage(mdlog, printHostName, useGpuForBonded, pmeRunMode, useGpuForUpdate);
1251 const bool disableNonbondedCalculation = (getenv("GMX_NO_NONBONDED") != nullptr);
1252 if (disableNonbondedCalculation)
1254 /* turn off non-bonded calculations */
1255 GMX_LOG(mdlog.warning)
1258 "Found environment variable GMX_NO_NONBONDED.\n"
1259 "Disabling nonbonded calculations.");
1262 MdrunScheduleWorkload runScheduleWork;
1263 // Also populates the simulation constant workload description.
1264 runScheduleWork.simulationWork =
1265 createSimulationWorkload(*inputrec, disableNonbondedCalculation, devFlags,
1266 useGpuForNonbonded, pmeRunMode, useGpuForBonded, useGpuForUpdate);
1268 std::unique_ptr<DeviceStreamManager> deviceStreamManager = nullptr;
1270 if (deviceInfo != nullptr)
1272 if (DOMAINDECOMP(cr) && thisRankHasDuty(cr, DUTY_PP))
1274 dd_setup_dlb_resource_sharing(cr, deviceId);
1276 deviceStreamManager = std::make_unique<DeviceStreamManager>(
1277 *deviceInfo, havePPDomainDecomposition(cr), runScheduleWork.simulationWork, useTiming);
1280 // If the user chose a task assignment, give them some hints
1281 // where appropriate.
1282 if (!userGpuTaskAssignment.empty())
1284 gpuTaskAssignments.logPerformanceHints(mdlog, ssize(gpuIdsToUse));
1289 /* After possible communicator splitting in make_dd_communicators.
1290 * we can set up the intra/inter node communication.
1292 gmx_setup_nodecomm(fplog, cr);
1298 GMX_LOG(mdlog.warning)
1300 .appendTextFormatted(
1301 "This is simulation %d out of %d running as a composite GROMACS\n"
1302 "multi-simulation job. Setup for this simulation:\n",
1303 ms->simulationIndex_, ms->numSimulations_);
1305 GMX_LOG(mdlog.warning)
1306 .appendTextFormatted("Using %d MPI %s\n", cr->nnodes,
1308 cr->nnodes == 1 ? "thread" : "threads"
1310 cr->nnodes == 1 ? "process" : "processes"
1316 // If mdrun -pin auto honors any affinity setting that already
1317 // exists. If so, it is nice to provide feedback about whether
1318 // that existing affinity setting was from OpenMP or something
1319 // else, so we run this code both before and after we initialize
1320 // the OpenMP support.
1321 gmx_check_thread_affinity_set(mdlog, &hw_opt, hwinfo->nthreads_hw_avail, FALSE);
1322 /* Check and update the number of OpenMP threads requested */
1323 checkAndUpdateRequestedNumOpenmpThreads(&hw_opt, *hwinfo, cr, ms, physicalNodeComm.size_,
1324 pmeRunMode, mtop, *inputrec);
1326 gmx_omp_nthreads_init(mdlog, cr, hwinfo->nthreads_hw_avail, physicalNodeComm.size_,
1327 hw_opt.nthreads_omp, hw_opt.nthreads_omp_pme, !thisRankHasDuty(cr, DUTY_PP));
1329 // Enable FP exception detection, but not in
1330 // Release mode and not for compilers with known buggy FP
1331 // exception support (clang with any optimization) or suspected
1332 // buggy FP exception support (gcc 7.* with optimization).
1333 #if !defined NDEBUG \
1334 && !((defined __clang__ || (defined(__GNUC__) && !defined(__ICC) && __GNUC__ == 7)) \
1335 && defined __OPTIMIZE__)
1336 const bool bEnableFPE = true;
1338 const bool bEnableFPE = false;
1340 // FIXME - reconcile with gmx_feenableexcept() call from CommandLineModuleManager::run()
1343 gmx_feenableexcept();
1346 /* Now that we know the setup is consistent, check for efficiency */
1347 check_resource_division_efficiency(hwinfo, gpuTaskAssignments.thisRankHasAnyGpuTask(),
1348 mdrunOptions.ntompOptionIsSet, cr, mdlog);
1350 /* getting number of PP/PME threads on this MPI / tMPI rank.
1351 PME: env variable should be read only on one node to make sure it is
1352 identical everywhere;
1354 const int numThreadsOnThisRank = thisRankHasDuty(cr, DUTY_PP) ? gmx_omp_nthreads_get(emntNonbonded)
1355 : gmx_omp_nthreads_get(emntPME);
1356 checkHardwareOversubscription(numThreadsOnThisRank, cr->nodeid, *hwinfo->hardwareTopology,
1357 physicalNodeComm, mdlog);
1359 // Enable Peer access between GPUs where available
1360 // Only for DD, only master PP rank needs to perform setup, and only if thread MPI plus
1361 // any of the GPU communication features are active.
1362 if (DOMAINDECOMP(cr) && MASTER(cr) && thisRankHasDuty(cr, DUTY_PP) && GMX_THREAD_MPI
1363 && (runScheduleWork.simulationWork.useGpuHaloExchange
1364 || runScheduleWork.simulationWork.useGpuPmePpCommunication))
1366 setupGpuDevicePeerAccess(gpuIdsToUse, mdlog);
1369 if (hw_opt.threadAffinity != ThreadAffinity::Off)
1371 /* Before setting affinity, check whether the affinity has changed
1372 * - which indicates that probably the OpenMP library has changed it
1373 * since we first checked).
1375 gmx_check_thread_affinity_set(mdlog, &hw_opt, hwinfo->nthreads_hw_avail, TRUE);
1377 int numThreadsOnThisNode, intraNodeThreadOffset;
1378 analyzeThreadsOnThisNode(physicalNodeComm, numThreadsOnThisRank, &numThreadsOnThisNode,
1379 &intraNodeThreadOffset);
1381 /* Set the CPU affinity */
1382 gmx_set_thread_affinity(mdlog, cr, &hw_opt, *hwinfo->hardwareTopology, numThreadsOnThisRank,
1383 numThreadsOnThisNode, intraNodeThreadOffset, nullptr);
1386 if (mdrunOptions.timingOptions.resetStep > -1)
1391 "The -resetstep functionality is deprecated, and may be removed in a "
1394 wcycle = wallcycle_init(fplog, mdrunOptions.timingOptions.resetStep, cr);
1398 /* Master synchronizes its value of reset_counters with all nodes
1399 * including PME only nodes */
1400 int64_t reset_counters = wcycle_get_reset_counters(wcycle);
1401 gmx_bcast(sizeof(reset_counters), &reset_counters, cr->mpi_comm_mysim);
1402 wcycle_set_reset_counters(wcycle, reset_counters);
1405 // Membrane embedding must be initialized before we call init_forcerec()
1406 membedHolder.initializeMembed(fplog, filenames.size(), filenames.data(), &mtop, inputrec.get(),
1407 globalState.get(), cr, &mdrunOptions.checkpointOptions.period);
1409 const bool thisRankHasPmeGpuTask = gpuTaskAssignments.thisRankHasPmeGpuTask();
1410 std::unique_ptr<MDAtoms> mdAtoms;
1411 std::unique_ptr<VirtualSitesHandler> vsite;
1412 std::unique_ptr<GpuBonded> gpuBonded;
1415 if (thisRankHasDuty(cr, DUTY_PP))
1417 mdModulesNotifier.notify(*cr);
1418 mdModulesNotifier.notify(&atomSets);
1419 mdModulesNotifier.notify(inputrec->pbcType);
1420 mdModulesNotifier.notify(SimulationTimeStep{ inputrec->delta_t });
1421 /* Initiate forcerecord */
1422 fr = new t_forcerec;
1423 fr->forceProviders = mdModules_->initForceProviders();
1424 init_forcerec(fplog, mdlog, fr, inputrec.get(), &mtop, cr, box,
1425 opt2fn("-table", filenames.size(), filenames.data()),
1426 opt2fn("-tablep", filenames.size(), filenames.data()),
1427 opt2fns("-tableb", filenames.size(), filenames.data()), pforce);
1428 // Dirty hack, for fixing disres and orires should be made mdmodules
1429 fr->fcdata->disres = disresdata;
1430 fr->fcdata->orires = oriresdata;
1432 // Save a handle to device stream manager to use elsewhere in the code
1433 // TODO: Forcerec is not a correct place to store it.
1434 fr->deviceStreamManager = deviceStreamManager.get();
1436 if (runScheduleWork.simulationWork.useGpuPmePpCommunication && !thisRankHasDuty(cr, DUTY_PME))
1439 deviceStreamManager != nullptr,
1440 "GPU device stream manager should be valid in order to use PME-PP direct "
1443 deviceStreamManager->streamIsValid(DeviceStreamType::PmePpTransfer),
1444 "GPU PP-PME stream should be valid in order to use GPU PME-PP direct "
1446 fr->pmePpCommGpu = std::make_unique<gmx::PmePpCommGpu>(
1447 cr->mpi_comm_mysim, cr->dd->pme_nodeid, deviceStreamManager->context(),
1448 deviceStreamManager->stream(DeviceStreamType::PmePpTransfer));
1451 fr->nbv = Nbnxm::init_nb_verlet(mdlog, inputrec.get(), fr, cr, *hwinfo,
1452 runScheduleWork.simulationWork.useGpuNonbonded,
1453 deviceStreamManager.get(), &mtop, box, wcycle);
1454 // TODO: Move the logic below to a GPU bonded builder
1455 if (runScheduleWork.simulationWork.useGpuBonded)
1457 GMX_RELEASE_ASSERT(deviceStreamManager != nullptr,
1458 "GPU device stream manager should be valid in order to use GPU "
1459 "version of bonded forces.");
1460 gpuBonded = std::make_unique<GpuBonded>(
1461 mtop.ffparams, fr->ic->epsfac * fr->fudgeQQ, deviceStreamManager->context(),
1462 deviceStreamManager->bondedStream(havePPDomainDecomposition(cr)), wcycle);
1463 fr->gpuBonded = gpuBonded.get();
1466 /* Initialize the mdAtoms structure.
1467 * mdAtoms is not filled with atom data,
1468 * as this can not be done now with domain decomposition.
1470 mdAtoms = makeMDAtoms(fplog, mtop, *inputrec, thisRankHasPmeGpuTask);
1471 if (globalState && thisRankHasPmeGpuTask)
1473 // The pinning of coordinates in the global state object works, because we only use
1474 // PME on GPU without DD or on a separate PME rank, and because the local state pointer
1475 // points to the global state object without DD.
1476 // FIXME: MD and EM separately set up the local state - this should happen in the same
1477 // function, which should also perform the pinning.
1478 changePinningPolicy(&globalState->x, pme_get_pinning_policy());
1481 /* Initialize the virtual site communication */
1482 vsite = makeVirtualSitesHandler(mtop, cr, fr->pbcType);
1484 calc_shifts(box, fr->shift_vec);
1486 /* With periodic molecules the charge groups should be whole at start up
1487 * and the virtual sites should not be far from their proper positions.
1489 if (!inputrec->bContinuation && MASTER(cr)
1490 && !(inputrec->pbcType != PbcType::No && inputrec->bPeriodicMols))
1492 /* Make molecules whole at start of run */
1493 if (fr->pbcType != PbcType::No)
1495 do_pbc_first_mtop(fplog, inputrec->pbcType, box, &mtop, globalState->x.rvec_array());
1499 /* Correct initial vsite positions are required
1500 * for the initial distribution in the domain decomposition
1501 * and for the initial shell prediction.
1503 constructVirtualSitesGlobal(mtop, globalState->x);
1507 if (EEL_PME(fr->ic->eeltype) || EVDW_PME(fr->ic->vdwtype))
1509 ewaldcoeff_q = fr->ic->ewaldcoeff_q;
1510 ewaldcoeff_lj = fr->ic->ewaldcoeff_lj;
1515 /* This is a PME only node */
1517 GMX_ASSERT(globalState == nullptr,
1518 "We don't need the state on a PME only rank and expect it to be unitialized");
1520 ewaldcoeff_q = calc_ewaldcoeff_q(inputrec->rcoulomb, inputrec->ewald_rtol);
1521 ewaldcoeff_lj = calc_ewaldcoeff_lj(inputrec->rvdw, inputrec->ewald_rtol_lj);
1524 gmx_pme_t* sepPmeData = nullptr;
1525 // This reference hides the fact that PME data is owned by runner on PME-only ranks and by forcerec on other ranks
1526 GMX_ASSERT(thisRankHasDuty(cr, DUTY_PP) == (fr != nullptr),
1527 "Double-checking that only PME-only ranks have no forcerec");
1528 gmx_pme_t*& pmedata = fr ? fr->pmedata : sepPmeData;
1530 // TODO should live in ewald module once its testing is improved
1532 // Later, this program could contain kernels that might be later
1533 // re-used as auto-tuning progresses, or subsequent simulations
1535 PmeGpuProgramStorage pmeGpuProgram;
1536 if (thisRankHasPmeGpuTask)
1539 (deviceStreamManager != nullptr),
1540 "GPU device stream manager should be initialized in order to use GPU for PME.");
1541 GMX_RELEASE_ASSERT((deviceInfo != nullptr),
1542 "GPU device should be initialized in order to use GPU for PME.");
1543 pmeGpuProgram = buildPmeGpuProgram(deviceStreamManager->context());
1546 /* Initiate PME if necessary,
1547 * either on all nodes or on dedicated PME nodes only. */
1548 if (EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype))
1550 if (mdAtoms && mdAtoms->mdatoms())
1552 nChargePerturbed = mdAtoms->mdatoms()->nChargePerturbed;
1553 if (EVDW_PME(inputrec->vdwtype))
1555 nTypePerturbed = mdAtoms->mdatoms()->nTypePerturbed;
1558 if (cr->npmenodes > 0)
1560 /* The PME only nodes need to know nChargePerturbed(FEP on Q) and nTypePerturbed(FEP on LJ)*/
1561 gmx_bcast(sizeof(nChargePerturbed), &nChargePerturbed, cr->mpi_comm_mysim);
1562 gmx_bcast(sizeof(nTypePerturbed), &nTypePerturbed, cr->mpi_comm_mysim);
1565 if (thisRankHasDuty(cr, DUTY_PME))
1569 // TODO: This should be in the builder.
1570 GMX_RELEASE_ASSERT(!runScheduleWork.simulationWork.useGpuPme
1571 || (deviceStreamManager != nullptr),
1572 "Device stream manager should be valid in order to use GPU "
1575 !runScheduleWork.simulationWork.useGpuPme
1576 || deviceStreamManager->streamIsValid(DeviceStreamType::Pme),
1577 "GPU PME stream should be valid in order to use GPU version of PME.");
1579 const DeviceContext* deviceContext = runScheduleWork.simulationWork.useGpuPme
1580 ? &deviceStreamManager->context()
1582 const DeviceStream* pmeStream =
1583 runScheduleWork.simulationWork.useGpuPme
1584 ? &deviceStreamManager->stream(DeviceStreamType::Pme)
1587 pmedata = gmx_pme_init(cr, getNumPmeDomains(cr->dd), inputrec.get(),
1588 nChargePerturbed != 0, nTypePerturbed != 0,
1589 mdrunOptions.reproducible, ewaldcoeff_q, ewaldcoeff_lj,
1590 gmx_omp_nthreads_get(emntPME), pmeRunMode, nullptr,
1591 deviceContext, pmeStream, pmeGpuProgram.get(), mdlog);
1593 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1598 if (EI_DYNAMICS(inputrec->eI))
1600 /* Turn on signal handling on all nodes */
1602 * (A user signal from the PME nodes (if any)
1603 * is communicated to the PP nodes.
1605 signal_handler_install();
1608 pull_t* pull_work = nullptr;
1609 if (thisRankHasDuty(cr, DUTY_PP))
1611 /* Assumes uniform use of the number of OpenMP threads */
1612 walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(emntDefault));
1614 if (inputrec->bPull)
1616 /* Initialize pull code */
1617 pull_work = init_pull(fplog, inputrec->pull, inputrec.get(), &mtop, cr, &atomSets,
1618 inputrec->fepvals->init_lambda);
1619 if (inputrec->pull->bXOutAverage || inputrec->pull->bFOutAverage)
1621 initPullHistory(pull_work, &observablesHistory);
1623 if (EI_DYNAMICS(inputrec->eI) && MASTER(cr))
1625 init_pull_output_files(pull_work, filenames.size(), filenames.data(), oenv, startingBehavior);
1629 std::unique_ptr<EnforcedRotation> enforcedRotation;
1632 /* Initialize enforced rotation code */
1633 enforcedRotation = init_rot(fplog, inputrec.get(), filenames.size(), filenames.data(),
1634 cr, &atomSets, globalState.get(), &mtop, oenv, mdrunOptions,
1638 t_swap* swap = nullptr;
1639 if (inputrec->eSwapCoords != eswapNO)
1641 /* Initialize ion swapping code */
1642 swap = init_swapcoords(fplog, inputrec.get(),
1643 opt2fn_master("-swap", filenames.size(), filenames.data(), cr),
1644 &mtop, globalState.get(), &observablesHistory, cr, &atomSets,
1645 oenv, mdrunOptions, startingBehavior);
1648 /* Let makeConstraints know whether we have essential dynamics constraints. */
1649 auto constr = makeConstraints(mtop, *inputrec, pull_work, doEssentialDynamics, fplog, cr,
1650 ms, &nrnb, wcycle, fr->bMolPBC);
1652 /* Energy terms and groups */
1653 gmx_enerdata_t enerd(mtop.groups.groups[SimulationAtomGroupType::EnergyOutput].size(),
1654 inputrec->fepvals->n_lambda);
1656 /* Kinetic energy data */
1657 gmx_ekindata_t ekind;
1658 init_ekindata(fplog, &mtop, &(inputrec->opts), &ekind, inputrec->cos_accel);
1660 /* Set up interactive MD (IMD) */
1662 makeImdSession(inputrec.get(), cr, wcycle, &enerd, ms, &mtop, mdlog,
1663 MASTER(cr) ? globalState->x.rvec_array() : nullptr, filenames.size(),
1664 filenames.data(), oenv, mdrunOptions.imdOptions, startingBehavior);
1666 if (DOMAINDECOMP(cr))
1668 GMX_RELEASE_ASSERT(fr, "fr was NULL while cr->duty was DUTY_PP");
1669 /* This call is not included in init_domain_decomposition mainly
1670 * because fr->cginfo_mb is set later.
1672 dd_init_bondeds(fplog, cr->dd, mtop, vsite.get(), inputrec.get(),
1673 domdecOptions.checkBondedInteractions, fr->cginfo_mb);
1676 if (runScheduleWork.simulationWork.useGpuBufferOps)
1678 fr->gpuForceReduction[gmx::AtomLocality::Local] = std::make_unique<gmx::GpuForceReduction>(
1679 deviceStreamManager->context(),
1680 deviceStreamManager->stream(gmx::DeviceStreamType::NonBondedLocal));
1681 fr->gpuForceReduction[gmx::AtomLocality::NonLocal] = std::make_unique<gmx::GpuForceReduction>(
1682 deviceStreamManager->context(),
1683 deviceStreamManager->stream(gmx::DeviceStreamType::NonBondedNonLocal));
1686 std::unique_ptr<gmx::StatePropagatorDataGpu> stateGpu;
1687 if (gpusWereDetected
1688 && ((runScheduleWork.simulationWork.useGpuPme && thisRankHasDuty(cr, DUTY_PME))
1689 || runScheduleWork.simulationWork.useGpuBufferOps))
1691 GpuApiCallBehavior transferKind = (inputrec->eI == eiMD && !doRerun && !useModularSimulator)
1692 ? GpuApiCallBehavior::Async
1693 : GpuApiCallBehavior::Sync;
1694 GMX_RELEASE_ASSERT(deviceStreamManager != nullptr,
1695 "GPU device stream manager should be initialized to use GPU.");
1696 stateGpu = std::make_unique<gmx::StatePropagatorDataGpu>(
1697 *deviceStreamManager, transferKind, pme_gpu_get_block_size(fr->pmedata), wcycle);
1698 fr->stateGpu = stateGpu.get();
1701 GMX_ASSERT(stopHandlerBuilder_, "Runner must provide StopHandlerBuilder to simulator.");
1702 SimulatorBuilder simulatorBuilder;
1704 simulatorBuilder.add(SimulatorStateData(globalState.get(), &observablesHistory, &enerd, &ekind));
1705 simulatorBuilder.add(std::move(membedHolder));
1706 simulatorBuilder.add(std::move(stopHandlerBuilder_));
1707 simulatorBuilder.add(SimulatorConfig(mdrunOptions, startingBehavior, &runScheduleWork));
1710 simulatorBuilder.add(SimulatorEnv(fplog, cr, ms, mdlog, oenv));
1711 simulatorBuilder.add(Profiling(&nrnb, walltime_accounting, wcycle));
1712 simulatorBuilder.add(ConstraintsParam(
1713 constr.get(), enforcedRotation ? enforcedRotation->getLegacyEnfrot() : nullptr,
1715 // TODO: Separate `fr` to a separate add, and make the `build` handle the coupling sensibly.
1716 simulatorBuilder.add(LegacyInput(static_cast<int>(filenames.size()), filenames.data(),
1717 inputrec.get(), fr));
1718 simulatorBuilder.add(ReplicaExchangeParameters(replExParams));
1719 simulatorBuilder.add(InteractiveMD(imdSession.get()));
1720 simulatorBuilder.add(SimulatorModules(mdModules_->outputProvider(), mdModules_->notifier()));
1721 simulatorBuilder.add(CenterOfMassPulling(pull_work));
1722 // Todo move to an MDModule
1723 simulatorBuilder.add(IonSwapping(swap));
1724 simulatorBuilder.add(TopologyData(&mtop, mdAtoms.get()));
1725 simulatorBuilder.add(BoxDeformationHandle(deform.get()));
1726 simulatorBuilder.add(std::move(modularSimulatorCheckpointData));
1728 // build and run simulator object based on user-input
1729 auto simulator = simulatorBuilder.build(useModularSimulator);
1732 if (fr->pmePpCommGpu)
1734 // destroy object since it is no longer required. (This needs to be done while the GPU context still exists.)
1735 fr->pmePpCommGpu.reset();
1738 if (inputrec->bPull)
1740 finish_pull(pull_work);
1742 finish_swapcoords(swap);
1746 GMX_RELEASE_ASSERT(pmedata, "pmedata was NULL while cr->duty was not DUTY_PP");
1748 walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(emntPME));
1749 gmx_pmeonly(pmedata, cr, &nrnb, wcycle, walltime_accounting, inputrec.get(), pmeRunMode,
1750 deviceStreamManager.get());
1753 wallcycle_stop(wcycle, ewcRUN);
1755 /* Finish up, write some stuff
1756 * if rerunMD, don't write last frame again
1758 finish_run(fplog, mdlog, cr, inputrec.get(), &nrnb, wcycle, walltime_accounting,
1759 fr ? fr->nbv.get() : nullptr, pmedata, EI_DYNAMICS(inputrec->eI) && !isMultiSim(ms));
1761 // clean up cycle counter
1762 wallcycle_destroy(wcycle);
1764 deviceStreamManager.reset(nullptr);
1768 gmx_pme_destroy(pmedata);
1772 // FIXME: this is only here to manually unpin mdAtoms->chargeA_ and state->x,
1773 // before we destroy the GPU context(s)
1774 // Pinned buffers are associated with contexts in CUDA.
1775 // As soon as we destroy GPU contexts after mdrunner() exits, these lines should go.
1776 mdAtoms.reset(nullptr);
1777 globalState.reset(nullptr);
1778 mdModules_.reset(nullptr); // destruct force providers here as they might also use the GPU
1779 gpuBonded.reset(nullptr);
1780 /* Free pinned buffers in *fr */
1783 // TODO convert to C++ so we can get rid of these frees
1787 if (!hwinfo->deviceInfoList.empty())
1789 /* stop the GPU profiler (only CUDA) */
1793 /* With tMPI we need to wait for all ranks to finish deallocation before
1794 * destroying the CUDA context as some tMPI ranks may be sharing
1797 * This is not a concern in OpenCL where we use one context per rank.
1799 * Note: it is safe to not call the barrier on the ranks which do not use GPU,
1800 * but it is easier and more futureproof to call it on the whole node.
1802 * Note that this function needs to be called even if GPUs are not used
1803 * in this run because the PME ranks have no knowledge of whether GPUs
1804 * are used or not, but all ranks need to enter the barrier below.
1805 * \todo Remove this physical node barrier after making sure
1806 * that it's not needed anymore (with a shared GPU run).
1810 physicalNodeComm.barrier();
1812 releaseDevice(deviceInfo);
1814 /* Does what it says */
1815 print_date_and_time(fplog, cr->nodeid, "Finished mdrun", gmx_gettime());
1816 walltime_accounting_destroy(walltime_accounting);
1818 // Ensure log file content is written
1821 gmx_fio_flush(logFileHandle);
1824 /* Reset FPEs (important for unit tests) by disabling them. Assumes no
1825 * exceptions were enabled before function was called. */
1828 gmx_fedisableexcept();
1831 auto rc = static_cast<int>(gmx_get_stop_condition());
1834 /* we need to join all threads. The sub-threads join when they
1835 exit this function, but the master thread needs to be told to
1845 Mdrunner::~Mdrunner()
1847 // Clean up of the Manager.
1848 // This will end up getting called on every thread-MPI rank, which is unnecessary,
1849 // but okay as long as threads synchronize some time before adding or accessing
1850 // a new set of restraints.
1851 if (restraintManager_)
1853 restraintManager_->clear();
1854 GMX_ASSERT(restraintManager_->countRestraints() == 0,
1855 "restraints added during runner life time should be cleared at runner "
1860 void Mdrunner::addPotential(std::shared_ptr<gmx::IRestraintPotential> puller, const std::string& name)
1862 GMX_ASSERT(restraintManager_, "Mdrunner must have a restraint manager.");
1863 // Not sure if this should be logged through the md logger or something else,
1864 // but it is helpful to have some sort of INFO level message sent somewhere.
1865 // std::cout << "Registering restraint named " << name << std::endl;
1867 // When multiple restraints are used, it may be wasteful to register them separately.
1868 // Maybe instead register an entire Restraint Manager as a force provider.
1869 restraintManager_->addToSpec(std::move(puller), name);
1872 Mdrunner::Mdrunner(std::unique_ptr<MDModules> mdModules) : mdModules_(std::move(mdModules)) {}
1874 Mdrunner::Mdrunner(Mdrunner&&) noexcept = default;
1876 Mdrunner& Mdrunner::operator=(Mdrunner&& /*handle*/) noexcept = default;
1878 class Mdrunner::BuilderImplementation
1881 BuilderImplementation() = delete;
1882 BuilderImplementation(std::unique_ptr<MDModules> mdModules, compat::not_null<SimulationContext*> context);
1883 ~BuilderImplementation();
1885 BuilderImplementation& setExtraMdrunOptions(const MdrunOptions& options,
1886 real forceWarningThreshold,
1887 StartingBehavior startingBehavior);
1889 void addDomdec(const DomdecOptions& options);
1891 void addInput(SimulationInputHandle inputHolder);
1893 void addVerletList(int nstlist);
1895 void addReplicaExchange(const ReplicaExchangeParameters& params);
1897 void addNonBonded(const char* nbpu_opt);
1899 void addPME(const char* pme_opt_, const char* pme_fft_opt_);
1901 void addBondedTaskAssignment(const char* bonded_opt);
1903 void addUpdateTaskAssignment(const char* update_opt);
1905 void addHardwareOptions(const gmx_hw_opt_t& hardwareOptions);
1907 void addFilenames(ArrayRef<const t_filenm> filenames);
1909 void addOutputEnvironment(gmx_output_env_t* outputEnvironment);
1911 void addLogFile(t_fileio* logFileHandle);
1913 void addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder);
1918 // Default parameters copied from runner.h
1919 // \todo Clarify source(s) of default parameters.
1921 const char* nbpu_opt_ = nullptr;
1922 const char* pme_opt_ = nullptr;
1923 const char* pme_fft_opt_ = nullptr;
1924 const char* bonded_opt_ = nullptr;
1925 const char* update_opt_ = nullptr;
1927 MdrunOptions mdrunOptions_;
1929 DomdecOptions domdecOptions_;
1931 ReplicaExchangeParameters replicaExchangeParameters_;
1933 //! Command-line override for the duration of a neighbor list with the Verlet scheme.
1936 //! World communicator, used for hardware detection and task assignment
1937 MPI_Comm libraryWorldCommunicator_ = MPI_COMM_NULL;
1939 //! Multisim communicator handle.
1940 gmx_multisim_t* multiSimulation_;
1942 //! mdrun communicator
1943 MPI_Comm simulationCommunicator_ = MPI_COMM_NULL;
1945 //! Print a warning if any force is larger than this (in kJ/mol nm).
1946 real forceWarningThreshold_ = -1;
1948 //! Whether the simulation will start afresh, or restart with/without appending.
1949 StartingBehavior startingBehavior_ = StartingBehavior::NewSimulation;
1951 //! The modules that comprise the functionality of mdrun.
1952 std::unique_ptr<MDModules> mdModules_;
1954 //! \brief Parallelism information.
1955 gmx_hw_opt_t hardwareOptions_;
1957 //! filename options for simulation.
1958 ArrayRef<const t_filenm> filenames_;
1960 /*! \brief Handle to output environment.
1962 * \todo gmx_output_env_t needs lifetime management.
1964 gmx_output_env_t* outputEnvironment_ = nullptr;
1966 /*! \brief Non-owning handle to MD log file.
1968 * \todo Context should own output facilities for client.
1969 * \todo Improve log file handle management.
1971 * Code managing the FILE* relies on the ability to set it to
1972 * nullptr to check whether the filehandle is valid.
1974 t_fileio* logFileHandle_ = nullptr;
1977 * \brief Builder for simulation stop signal handler.
1979 std::unique_ptr<StopHandlerBuilder> stopHandlerBuilder_ = nullptr;
1982 * \brief Sources for initial simulation state.
1984 * See issue #3652 for near-term refinements to the SimulationInput interface.
1986 * See issue #3379 for broader discussion on API aspects of simulation inputs and outputs.
1988 SimulationInputHandle inputHolder_;
1991 Mdrunner::BuilderImplementation::BuilderImplementation(std::unique_ptr<MDModules> mdModules,
1992 compat::not_null<SimulationContext*> context) :
1993 mdModules_(std::move(mdModules))
1995 libraryWorldCommunicator_ = context->libraryWorldCommunicator_;
1996 simulationCommunicator_ = context->simulationCommunicator_;
1997 multiSimulation_ = context->multiSimulation_.get();
2000 Mdrunner::BuilderImplementation::~BuilderImplementation() = default;
2002 Mdrunner::BuilderImplementation&
2003 Mdrunner::BuilderImplementation::setExtraMdrunOptions(const MdrunOptions& options,
2004 const real forceWarningThreshold,
2005 const StartingBehavior startingBehavior)
2007 mdrunOptions_ = options;
2008 forceWarningThreshold_ = forceWarningThreshold;
2009 startingBehavior_ = startingBehavior;
2013 void Mdrunner::BuilderImplementation::addDomdec(const DomdecOptions& options)
2015 domdecOptions_ = options;
2018 void Mdrunner::BuilderImplementation::addVerletList(int nstlist)
2023 void Mdrunner::BuilderImplementation::addReplicaExchange(const ReplicaExchangeParameters& params)
2025 replicaExchangeParameters_ = params;
2028 Mdrunner Mdrunner::BuilderImplementation::build()
2030 auto newRunner = Mdrunner(std::move(mdModules_));
2032 newRunner.mdrunOptions = mdrunOptions_;
2033 newRunner.pforce = forceWarningThreshold_;
2034 newRunner.startingBehavior = startingBehavior_;
2035 newRunner.domdecOptions = domdecOptions_;
2037 // \todo determine an invariant to check or confirm that all gmx_hw_opt_t objects are valid
2038 newRunner.hw_opt = hardwareOptions_;
2040 // No invariant to check. This parameter exists to optionally override other behavior.
2041 newRunner.nstlist_cmdline = nstlist_;
2043 newRunner.replExParams = replicaExchangeParameters_;
2045 newRunner.filenames = filenames_;
2047 newRunner.libraryWorldCommunicator = libraryWorldCommunicator_;
2049 newRunner.simulationCommunicator = simulationCommunicator_;
2051 // nullptr is a valid value for the multisim handle
2052 newRunner.ms = multiSimulation_;
2056 newRunner.inputHolder_ = std::move(inputHolder_);
2060 GMX_THROW(gmx::APIError("MdrunnerBuilder::addInput() is required before build()."));
2063 // \todo Clarify ownership and lifetime management for gmx_output_env_t
2064 // \todo Update sanity checking when output environment has clearly specified invariants.
2065 // Initialization and default values for oenv are not well specified in the current version.
2066 if (outputEnvironment_)
2068 newRunner.oenv = outputEnvironment_;
2072 GMX_THROW(gmx::APIError(
2073 "MdrunnerBuilder::addOutputEnvironment() is required before build()"));
2076 newRunner.logFileHandle = logFileHandle_;
2080 newRunner.nbpu_opt = nbpu_opt_;
2084 GMX_THROW(gmx::APIError("MdrunnerBuilder::addNonBonded() is required before build()"));
2087 if (pme_opt_ && pme_fft_opt_)
2089 newRunner.pme_opt = pme_opt_;
2090 newRunner.pme_fft_opt = pme_fft_opt_;
2094 GMX_THROW(gmx::APIError("MdrunnerBuilder::addElectrostatics() is required before build()"));
2099 newRunner.bonded_opt = bonded_opt_;
2103 GMX_THROW(gmx::APIError(
2104 "MdrunnerBuilder::addBondedTaskAssignment() is required before build()"));
2109 newRunner.update_opt = update_opt_;
2113 GMX_THROW(gmx::APIError(
2114 "MdrunnerBuilder::addUpdateTaskAssignment() is required before build() "));
2118 newRunner.restraintManager_ = std::make_unique<gmx::RestraintManager>();
2120 if (stopHandlerBuilder_)
2122 newRunner.stopHandlerBuilder_ = std::move(stopHandlerBuilder_);
2126 newRunner.stopHandlerBuilder_ = std::make_unique<StopHandlerBuilder>();
2132 void Mdrunner::BuilderImplementation::addNonBonded(const char* nbpu_opt)
2134 nbpu_opt_ = nbpu_opt;
2137 void Mdrunner::BuilderImplementation::addPME(const char* pme_opt, const char* pme_fft_opt)
2140 pme_fft_opt_ = pme_fft_opt;
2143 void Mdrunner::BuilderImplementation::addBondedTaskAssignment(const char* bonded_opt)
2145 bonded_opt_ = bonded_opt;
2148 void Mdrunner::BuilderImplementation::addUpdateTaskAssignment(const char* update_opt)
2150 update_opt_ = update_opt;
2153 void Mdrunner::BuilderImplementation::addHardwareOptions(const gmx_hw_opt_t& hardwareOptions)
2155 hardwareOptions_ = hardwareOptions;
2158 void Mdrunner::BuilderImplementation::addFilenames(ArrayRef<const t_filenm> filenames)
2160 filenames_ = filenames;
2163 void Mdrunner::BuilderImplementation::addOutputEnvironment(gmx_output_env_t* outputEnvironment)
2165 outputEnvironment_ = outputEnvironment;
2168 void Mdrunner::BuilderImplementation::addLogFile(t_fileio* logFileHandle)
2170 logFileHandle_ = logFileHandle;
2173 void Mdrunner::BuilderImplementation::addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder)
2175 stopHandlerBuilder_ = std::move(builder);
2178 void Mdrunner::BuilderImplementation::addInput(SimulationInputHandle inputHolder)
2180 inputHolder_ = std::move(inputHolder);
2183 MdrunnerBuilder::MdrunnerBuilder(std::unique_ptr<MDModules> mdModules,
2184 compat::not_null<SimulationContext*> context) :
2185 impl_{ std::make_unique<Mdrunner::BuilderImplementation>(std::move(mdModules), context) }
2189 MdrunnerBuilder::~MdrunnerBuilder() = default;
2191 MdrunnerBuilder& MdrunnerBuilder::addSimulationMethod(const MdrunOptions& options,
2192 real forceWarningThreshold,
2193 const StartingBehavior startingBehavior)
2195 impl_->setExtraMdrunOptions(options, forceWarningThreshold, startingBehavior);
2199 MdrunnerBuilder& MdrunnerBuilder::addDomainDecomposition(const DomdecOptions& options)
2201 impl_->addDomdec(options);
2205 MdrunnerBuilder& MdrunnerBuilder::addNeighborList(int nstlist)
2207 impl_->addVerletList(nstlist);
2211 MdrunnerBuilder& MdrunnerBuilder::addReplicaExchange(const ReplicaExchangeParameters& params)
2213 impl_->addReplicaExchange(params);
2217 MdrunnerBuilder& MdrunnerBuilder::addNonBonded(const char* nbpu_opt)
2219 impl_->addNonBonded(nbpu_opt);
2223 MdrunnerBuilder& MdrunnerBuilder::addElectrostatics(const char* pme_opt, const char* pme_fft_opt)
2225 // The builder method may become more general in the future, but in this version,
2226 // parameters for PME electrostatics are both required and the only parameters
2228 if (pme_opt && pme_fft_opt)
2230 impl_->addPME(pme_opt, pme_fft_opt);
2235 gmx::InvalidInputError("addElectrostatics() arguments must be non-null pointers."));
2240 MdrunnerBuilder& MdrunnerBuilder::addBondedTaskAssignment(const char* bonded_opt)
2242 impl_->addBondedTaskAssignment(bonded_opt);
2246 MdrunnerBuilder& MdrunnerBuilder::addUpdateTaskAssignment(const char* update_opt)
2248 impl_->addUpdateTaskAssignment(update_opt);
2252 Mdrunner MdrunnerBuilder::build()
2254 return impl_->build();
2257 MdrunnerBuilder& MdrunnerBuilder::addHardwareOptions(const gmx_hw_opt_t& hardwareOptions)
2259 impl_->addHardwareOptions(hardwareOptions);
2263 MdrunnerBuilder& MdrunnerBuilder::addFilenames(ArrayRef<const t_filenm> filenames)
2265 impl_->addFilenames(filenames);
2269 MdrunnerBuilder& MdrunnerBuilder::addOutputEnvironment(gmx_output_env_t* outputEnvironment)
2271 impl_->addOutputEnvironment(outputEnvironment);
2275 MdrunnerBuilder& MdrunnerBuilder::addLogFile(t_fileio* logFileHandle)
2277 impl_->addLogFile(logFileHandle);
2281 MdrunnerBuilder& MdrunnerBuilder::addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder)
2283 impl_->addStopHandlerBuilder(std::move(builder));
2287 MdrunnerBuilder& MdrunnerBuilder::addInput(SimulationInputHandle input)
2289 impl_->addInput(std::move(input));
2293 MdrunnerBuilder::MdrunnerBuilder(MdrunnerBuilder&&) noexcept = default;
2295 MdrunnerBuilder& MdrunnerBuilder::operator=(MdrunnerBuilder&&) noexcept = default;