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_stream_manager.h"
77 #include "gromacs/hardware/cpuinfo.h"
78 #include "gromacs/hardware/detecthardware.h"
79 #include "gromacs/hardware/device_management.h"
80 #include "gromacs/hardware/printhardware.h"
81 #include "gromacs/imd/imd.h"
82 #include "gromacs/listed_forces/disre.h"
83 #include "gromacs/listed_forces/gpubonded.h"
84 #include "gromacs/listed_forces/listed_forces.h"
85 #include "gromacs/listed_forces/orires.h"
86 #include "gromacs/math/functions.h"
87 #include "gromacs/math/utilities.h"
88 #include "gromacs/math/vec.h"
89 #include "gromacs/mdlib/boxdeformation.h"
90 #include "gromacs/mdlib/broadcaststructs.h"
91 #include "gromacs/mdlib/calc_verletbuf.h"
92 #include "gromacs/mdlib/dispersioncorrection.h"
93 #include "gromacs/mdlib/enerdata_utils.h"
94 #include "gromacs/mdlib/force.h"
95 #include "gromacs/mdlib/forcerec.h"
96 #include "gromacs/mdlib/gmx_omp_nthreads.h"
97 #include "gromacs/mdlib/gpuforcereduction.h"
98 #include "gromacs/mdlib/makeconstraints.h"
99 #include "gromacs/mdlib/md_support.h"
100 #include "gromacs/mdlib/mdatoms.h"
101 #include "gromacs/mdlib/sighandler.h"
102 #include "gromacs/mdlib/stophandler.h"
103 #include "gromacs/mdlib/tgroup.h"
104 #include "gromacs/mdlib/updategroups.h"
105 #include "gromacs/mdlib/vsite.h"
106 #include "gromacs/mdrun/mdmodules.h"
107 #include "gromacs/mdrun/simulationcontext.h"
108 #include "gromacs/mdrun/simulationinput.h"
109 #include "gromacs/mdrun/simulationinputhandle.h"
110 #include "gromacs/mdrunutility/handlerestart.h"
111 #include "gromacs/mdrunutility/logging.h"
112 #include "gromacs/mdrunutility/multisim.h"
113 #include "gromacs/mdrunutility/printtime.h"
114 #include "gromacs/mdrunutility/threadaffinity.h"
115 #include "gromacs/mdtypes/checkpointdata.h"
116 #include "gromacs/mdtypes/commrec.h"
117 #include "gromacs/mdtypes/enerdata.h"
118 #include "gromacs/mdtypes/fcdata.h"
119 #include "gromacs/mdtypes/forcerec.h"
120 #include "gromacs/mdtypes/group.h"
121 #include "gromacs/mdtypes/inputrec.h"
122 #include "gromacs/mdtypes/interaction_const.h"
123 #include "gromacs/mdtypes/md_enums.h"
124 #include "gromacs/mdtypes/mdatom.h"
125 #include "gromacs/mdtypes/mdrunoptions.h"
126 #include "gromacs/mdtypes/observableshistory.h"
127 #include "gromacs/mdtypes/simulation_workload.h"
128 #include "gromacs/mdtypes/state.h"
129 #include "gromacs/mdtypes/state_propagator_data_gpu.h"
130 #include "gromacs/modularsimulator/modularsimulator.h"
131 #include "gromacs/nbnxm/gpu_data_mgmt.h"
132 #include "gromacs/nbnxm/nbnxm.h"
133 #include "gromacs/nbnxm/pairlist_tuning.h"
134 #include "gromacs/pbcutil/pbc.h"
135 #include "gromacs/pulling/output.h"
136 #include "gromacs/pulling/pull.h"
137 #include "gromacs/pulling/pull_rotation.h"
138 #include "gromacs/restraint/manager.h"
139 #include "gromacs/restraint/restraintmdmodule.h"
140 #include "gromacs/restraint/restraintpotential.h"
141 #include "gromacs/swap/swapcoords.h"
142 #include "gromacs/taskassignment/decidegpuusage.h"
143 #include "gromacs/taskassignment/decidesimulationworkload.h"
144 #include "gromacs/taskassignment/resourcedivision.h"
145 #include "gromacs/taskassignment/taskassignment.h"
146 #include "gromacs/taskassignment/usergpuids.h"
147 #include "gromacs/timing/gpu_timing.h"
148 #include "gromacs/timing/wallcycle.h"
149 #include "gromacs/timing/wallcyclereporting.h"
150 #include "gromacs/topology/mtop_util.h"
151 #include "gromacs/trajectory/trajectoryframe.h"
152 #include "gromacs/utility/basenetwork.h"
153 #include "gromacs/utility/cstringutil.h"
154 #include "gromacs/utility/exceptions.h"
155 #include "gromacs/utility/fatalerror.h"
156 #include "gromacs/utility/filestream.h"
157 #include "gromacs/utility/gmxassert.h"
158 #include "gromacs/utility/gmxmpi.h"
159 #include "gromacs/utility/keyvaluetree.h"
160 #include "gromacs/utility/logger.h"
161 #include "gromacs/utility/loggerbuilder.h"
162 #include "gromacs/utility/mdmodulenotification.h"
163 #include "gromacs/utility/physicalnodecommunicator.h"
164 #include "gromacs/utility/pleasecite.h"
165 #include "gromacs/utility/programcontext.h"
166 #include "gromacs/utility/smalloc.h"
167 #include "gromacs/utility/stringutil.h"
169 #include "isimulator.h"
170 #include "membedholder.h"
171 #include "replicaexchange.h"
172 #include "simulatorbuilder.h"
175 # include "corewrap.h"
182 /*! \brief Manage any development feature flag variables encountered
184 * The use of dev features indicated by environment variables is
185 * logged in order to ensure that runs with such features enabled can
186 * be identified from their log and standard output. Any cross
187 * dependencies are also checked, and if unsatisfied, a fatal error
190 * Note that some development features overrides are applied already here:
191 * the GPU communication flags are set to false in non-tMPI and non-CUDA builds.
193 * \param[in] mdlog Logger object.
194 * \param[in] useGpuForNonbonded True if the nonbonded task is offloaded in this run.
195 * \param[in] pmeRunMode The PME run mode for this run
196 * \returns The object populated with development feature flags.
198 static DevelopmentFeatureFlags manageDevelopmentFeatures(const gmx::MDLogger& mdlog,
199 const bool useGpuForNonbonded,
200 const PmeRunMode pmeRunMode)
202 DevelopmentFeatureFlags devFlags;
204 // Some builds of GCC 5 give false positive warnings that these
205 // getenv results are ignored when clearly they are used.
206 #pragma GCC diagnostic push
207 #pragma GCC diagnostic ignored "-Wunused-result"
209 devFlags.enableGpuBufferOps =
210 GMX_GPU_CUDA && useGpuForNonbonded && (getenv("GMX_USE_GPU_BUFFER_OPS") != nullptr);
211 devFlags.enableGpuHaloExchange = GMX_GPU_CUDA && GMX_THREAD_MPI && getenv("GMX_GPU_DD_COMMS") != nullptr;
212 devFlags.forceGpuUpdateDefault = (getenv("GMX_FORCE_UPDATE_DEFAULT_GPU") != nullptr);
213 devFlags.enableGpuPmePPComm =
214 GMX_GPU_CUDA && GMX_THREAD_MPI && getenv("GMX_GPU_PME_PP_COMMS") != nullptr;
216 #pragma GCC diagnostic pop
218 if (devFlags.enableGpuBufferOps)
220 GMX_LOG(mdlog.warning)
222 .appendTextFormatted(
223 "This run uses the 'GPU buffer ops' feature, enabled by the "
224 "GMX_USE_GPU_BUFFER_OPS environment variable.");
227 if (devFlags.forceGpuUpdateDefault)
229 GMX_LOG(mdlog.warning)
231 .appendTextFormatted(
232 "This run will default to '-update gpu' as requested by the "
233 "GMX_FORCE_UPDATE_DEFAULT_GPU environment variable. GPU update with domain "
234 "decomposition lacks substantial testing and should be used with caution.");
237 if (devFlags.enableGpuHaloExchange)
239 if (useGpuForNonbonded)
241 if (!devFlags.enableGpuBufferOps)
243 GMX_LOG(mdlog.warning)
245 .appendTextFormatted(
246 "Enabling GPU buffer operations required by GMX_GPU_DD_COMMS "
247 "(equivalent with GMX_USE_GPU_BUFFER_OPS=1).");
248 devFlags.enableGpuBufferOps = true;
250 GMX_LOG(mdlog.warning)
252 .appendTextFormatted(
253 "This run has requested the 'GPU halo exchange' feature, enabled by "
255 "GMX_GPU_DD_COMMS environment variable.");
259 GMX_LOG(mdlog.warning)
261 .appendTextFormatted(
262 "GMX_GPU_DD_COMMS environment variable detected, but the 'GPU "
263 "halo exchange' feature will not be enabled as nonbonded interactions "
264 "are not offloaded.");
265 devFlags.enableGpuHaloExchange = false;
269 if (devFlags.enableGpuPmePPComm)
271 if (pmeRunMode == PmeRunMode::GPU)
273 if (!devFlags.enableGpuBufferOps)
275 GMX_LOG(mdlog.warning)
277 .appendTextFormatted(
278 "Enabling GPU buffer operations required by GMX_GPU_PME_PP_COMMS "
279 "(equivalent with GMX_USE_GPU_BUFFER_OPS=1).");
280 devFlags.enableGpuBufferOps = true;
282 GMX_LOG(mdlog.warning)
284 .appendTextFormatted(
285 "This run uses the 'GPU PME-PP communications' feature, enabled "
286 "by the GMX_GPU_PME_PP_COMMS environment variable.");
290 std::string clarification;
291 if (pmeRunMode == PmeRunMode::Mixed)
294 "PME FFT and gather are not offloaded to the GPU (PME is running in mixed "
299 clarification = "PME is not offloaded to the GPU.";
301 GMX_LOG(mdlog.warning)
304 "GMX_GPU_PME_PP_COMMS environment variable detected, but the "
305 "'GPU PME-PP communications' feature was not enabled as "
307 devFlags.enableGpuPmePPComm = false;
314 /*! \brief Barrier for safe simultaneous thread access to mdrunner data
316 * Used to ensure that the master thread does not modify mdrunner during copy
317 * on the spawned threads. */
318 static void threadMpiMdrunnerAccessBarrier()
321 MPI_Barrier(MPI_COMM_WORLD);
325 Mdrunner Mdrunner::cloneOnSpawnedThread() const
327 auto newRunner = Mdrunner(std::make_unique<MDModules>());
329 // All runners in the same process share a restraint manager resource because it is
330 // part of the interface to the client code, which is associated only with the
331 // original thread. Handles to the same resources can be obtained by copy.
333 newRunner.restraintManager_ = std::make_unique<RestraintManager>(*restraintManager_);
336 // Copy members of master runner.
337 // \todo Replace with builder when Simulation context and/or runner phases are better defined.
338 // Ref https://gitlab.com/gromacs/gromacs/-/issues/2587 and https://gitlab.com/gromacs/gromacs/-/issues/2375
339 newRunner.hw_opt = hw_opt;
340 newRunner.filenames = filenames;
342 newRunner.oenv = oenv;
343 newRunner.mdrunOptions = mdrunOptions;
344 newRunner.domdecOptions = domdecOptions;
345 newRunner.nbpu_opt = nbpu_opt;
346 newRunner.pme_opt = pme_opt;
347 newRunner.pme_fft_opt = pme_fft_opt;
348 newRunner.bonded_opt = bonded_opt;
349 newRunner.update_opt = update_opt;
350 newRunner.nstlist_cmdline = nstlist_cmdline;
351 newRunner.replExParams = replExParams;
352 newRunner.pforce = pforce;
353 // Give the spawned thread the newly created valid communicator
354 // for the simulation.
355 newRunner.libraryWorldCommunicator = MPI_COMM_WORLD;
356 newRunner.simulationCommunicator = MPI_COMM_WORLD;
358 newRunner.startingBehavior = startingBehavior;
359 newRunner.stopHandlerBuilder_ = std::make_unique<StopHandlerBuilder>(*stopHandlerBuilder_);
360 newRunner.inputHolder_ = inputHolder_;
362 threadMpiMdrunnerAccessBarrier();
367 /*! \brief The callback used for running on spawned threads.
369 * Obtains the pointer to the master mdrunner object from the one
370 * argument permitted to the thread-launch API call, copies it to make
371 * a new runner for this thread, reinitializes necessary data, and
372 * proceeds to the simulation. */
373 static void mdrunner_start_fn(const void* arg)
377 auto masterMdrunner = reinterpret_cast<const gmx::Mdrunner*>(arg);
378 /* copy the arg list to make sure that it's thread-local. This
379 doesn't copy pointed-to items, of course; fnm, cr and fplog
380 are reset in the call below, all others should be const. */
381 gmx::Mdrunner mdrunner = masterMdrunner->cloneOnSpawnedThread();
384 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
388 void Mdrunner::spawnThreads(int numThreadsToLaunch)
391 /* now spawn new threads that start mdrunner_start_fn(), while
392 the main thread returns. Thread affinity is handled later. */
393 if (tMPI_Init_fn(TRUE, numThreadsToLaunch, TMPI_AFFINITY_NONE, mdrunner_start_fn,
394 static_cast<const void*>(this))
397 GMX_THROW(gmx::InternalError("Failed to spawn thread-MPI threads"));
400 // Give the master thread the newly created valid communicator for
402 libraryWorldCommunicator = MPI_COMM_WORLD;
403 simulationCommunicator = MPI_COMM_WORLD;
404 threadMpiMdrunnerAccessBarrier();
406 GMX_UNUSED_VALUE(numThreadsToLaunch);
407 GMX_UNUSED_VALUE(mdrunner_start_fn);
413 /*! \brief Initialize variables for Verlet scheme simulation */
414 static void prepare_verlet_scheme(FILE* fplog,
418 const gmx_mtop_t* mtop,
420 bool makeGpuPairList,
421 const gmx::CpuInfo& cpuinfo)
423 // We checked the cut-offs in grompp, but double-check here.
424 // We have PME+LJcutoff kernels for rcoulomb>rvdw.
425 if (EEL_PME_EWALD(ir->coulombtype) && ir->vdwtype == eelCUT)
427 GMX_RELEASE_ASSERT(ir->rcoulomb >= ir->rvdw,
428 "With Verlet lists and PME we should have rcoulomb>=rvdw");
432 GMX_RELEASE_ASSERT(ir->rcoulomb == ir->rvdw,
433 "With Verlet lists and no PME rcoulomb and rvdw should be identical");
435 /* For NVE simulations, we will retain the initial list buffer */
436 if (EI_DYNAMICS(ir->eI) && ir->verletbuf_tol > 0 && !(EI_MD(ir->eI) && ir->etc == etcNO))
438 /* Update the Verlet buffer size for the current run setup */
440 /* Here we assume SIMD-enabled kernels are being used. But as currently
441 * calc_verlet_buffer_size gives the same results for 4x8 and 4x4
442 * and 4x2 gives a larger buffer than 4x4, this is ok.
444 ListSetupType listType =
445 (makeGpuPairList ? ListSetupType::Gpu : ListSetupType::CpuSimdWhenSupported);
446 VerletbufListSetup listSetup = verletbufGetSafeListSetup(listType);
448 const real rlist_new =
449 calcVerletBufferSize(*mtop, det(box), *ir, ir->nstlist, ir->nstlist - 1, -1, listSetup);
451 if (rlist_new != ir->rlist)
453 if (fplog != nullptr)
456 "\nChanging rlist from %g to %g for non-bonded %dx%d atom kernels\n\n",
457 ir->rlist, rlist_new, listSetup.cluster_size_i, listSetup.cluster_size_j);
459 ir->rlist = rlist_new;
463 if (nstlist_cmdline > 0 && (!EI_DYNAMICS(ir->eI) || ir->verletbuf_tol <= 0))
465 gmx_fatal(FARGS, "Can not set nstlist without %s",
466 !EI_DYNAMICS(ir->eI) ? "dynamics" : "verlet-buffer-tolerance");
469 if (EI_DYNAMICS(ir->eI))
471 /* Set or try nstlist values */
472 increaseNstlist(fplog, cr, ir, nstlist_cmdline, mtop, box, makeGpuPairList, cpuinfo);
476 /*! \brief Override the nslist value in inputrec
478 * with value passed on the command line (if any)
480 static void override_nsteps_cmdline(const gmx::MDLogger& mdlog, int64_t nsteps_cmdline, t_inputrec* ir)
484 /* override with anything else than the default -2 */
485 if (nsteps_cmdline > -2)
487 char sbuf_steps[STEPSTRSIZE];
488 char sbuf_msg[STRLEN];
490 ir->nsteps = nsteps_cmdline;
491 if (EI_DYNAMICS(ir->eI) && nsteps_cmdline != -1)
494 "Overriding nsteps with value passed on the command line: %s steps, %.3g ps",
495 gmx_step_str(nsteps_cmdline, sbuf_steps), fabs(nsteps_cmdline * ir->delta_t));
499 sprintf(sbuf_msg, "Overriding nsteps with value passed on the command line: %s steps",
500 gmx_step_str(nsteps_cmdline, sbuf_steps));
503 GMX_LOG(mdlog.warning).asParagraph().appendText(sbuf_msg);
505 else if (nsteps_cmdline < -2)
507 gmx_fatal(FARGS, "Invalid nsteps value passed on the command line: %" PRId64, nsteps_cmdline);
509 /* Do nothing if nsteps_cmdline == -2 */
515 /*! \brief Return whether GPU acceleration of nonbondeds is supported with the given settings.
517 * If not, and if a warning may be issued, logs a warning about
518 * falling back to CPU code. With thread-MPI, only the first
519 * call to this function should have \c issueWarning true. */
520 static bool gpuAccelerationOfNonbondedIsUseful(const MDLogger& mdlog, const t_inputrec& ir, bool issueWarning)
522 bool gpuIsUseful = true;
525 if (ir.opts.ngener - ir.nwall > 1)
527 /* The GPU code does not support more than one energy group.
528 * If the user requested GPUs explicitly, a fatal error is given later.
532 "Multiple energy groups is not implemented for GPUs, falling back to the CPU. "
533 "For better performance, run on the GPU without energy groups and then do "
534 "gmx mdrun -rerun option on the trajectory with an energy group .tpr file.";
540 warning = "TPI is not implemented for GPUs.";
543 if (!gpuIsUseful && issueWarning)
545 GMX_LOG(mdlog.warning).asParagraph().appendText(warning);
551 //! Initializes the logger for mdrun.
552 static gmx::LoggerOwner buildLogger(FILE* fplog, const bool isSimulationMasterRank)
554 gmx::LoggerBuilder builder;
555 if (fplog != nullptr)
557 builder.addTargetFile(gmx::MDLogger::LogLevel::Info, fplog);
559 if (isSimulationMasterRank)
561 builder.addTargetStream(gmx::MDLogger::LogLevel::Warning, &gmx::TextOutputFile::standardError());
563 return builder.build();
566 //! Make a TaskTarget from an mdrun argument string.
567 static TaskTarget findTaskTarget(const char* optionString)
569 TaskTarget returnValue = TaskTarget::Auto;
571 if (strncmp(optionString, "auto", 3) == 0)
573 returnValue = TaskTarget::Auto;
575 else if (strncmp(optionString, "cpu", 3) == 0)
577 returnValue = TaskTarget::Cpu;
579 else if (strncmp(optionString, "gpu", 3) == 0)
581 returnValue = TaskTarget::Gpu;
585 GMX_ASSERT(false, "Option string should have been checked for sanity already");
591 //! Finish run, aggregate data to print performance info.
592 static void finish_run(FILE* fplog,
593 const gmx::MDLogger& mdlog,
595 const t_inputrec* inputrec,
597 gmx_wallcycle_t wcycle,
598 gmx_walltime_accounting_t walltime_accounting,
599 nonbonded_verlet_t* nbv,
600 const gmx_pme_t* pme,
604 double nbfs = 0, mflop = 0;
605 double elapsed_time, elapsed_time_over_all_ranks, elapsed_time_over_all_threads,
606 elapsed_time_over_all_threads_over_all_ranks;
607 /* Control whether it is valid to print a report. Only the
608 simulation master may print, but it should not do so if the run
609 terminated e.g. before a scheduled reset step. This is
610 complicated by the fact that PME ranks are unaware of the
611 reason why they were sent a pmerecvqxFINISH. To avoid
612 communication deadlocks, we always do the communication for the
613 report, even if we've decided not to write the report, because
614 how long it takes to finish the run is not important when we've
615 decided not to report on the simulation performance.
617 Further, we only report performance for dynamical integrators,
618 because those are the only ones for which we plan to
619 consider doing any optimizations. */
620 bool printReport = EI_DYNAMICS(inputrec->eI) && SIMMASTER(cr);
622 if (printReport && !walltime_accounting_get_valid_finish(walltime_accounting))
624 GMX_LOG(mdlog.warning)
626 .appendText("Simulation ended prematurely, no performance report will be written.");
631 std::unique_ptr<t_nrnb> nrnbTotalStorage;
634 nrnbTotalStorage = std::make_unique<t_nrnb>();
635 nrnb_tot = nrnbTotalStorage.get();
637 MPI_Allreduce(nrnb->n, nrnb_tot->n, eNRNB, MPI_DOUBLE, MPI_SUM, cr->mpi_comm_mysim);
645 elapsed_time = walltime_accounting_get_time_since_reset(walltime_accounting);
646 elapsed_time_over_all_threads =
647 walltime_accounting_get_time_since_reset_over_all_threads(walltime_accounting);
651 /* reduce elapsed_time over all MPI ranks in the current simulation */
652 MPI_Allreduce(&elapsed_time, &elapsed_time_over_all_ranks, 1, MPI_DOUBLE, MPI_SUM,
654 elapsed_time_over_all_ranks /= cr->nnodes;
655 /* Reduce elapsed_time_over_all_threads over all MPI ranks in the
656 * current simulation. */
657 MPI_Allreduce(&elapsed_time_over_all_threads, &elapsed_time_over_all_threads_over_all_ranks,
658 1, MPI_DOUBLE, MPI_SUM, cr->mpi_comm_mysim);
663 elapsed_time_over_all_ranks = elapsed_time;
664 elapsed_time_over_all_threads_over_all_ranks = elapsed_time_over_all_threads;
669 print_flop(fplog, nrnb_tot, &nbfs, &mflop);
672 if (thisRankHasDuty(cr, DUTY_PP) && DOMAINDECOMP(cr))
674 print_dd_statistics(cr, inputrec, fplog);
677 /* TODO Move the responsibility for any scaling by thread counts
678 * to the code that handled the thread region, so that there's a
679 * mechanism to keep cycle counting working during the transition
680 * to task parallelism. */
681 int nthreads_pp = gmx_omp_nthreads_get(emntNonbonded);
682 int nthreads_pme = gmx_omp_nthreads_get(emntPME);
683 wallcycle_scale_by_num_threads(wcycle, thisRankHasDuty(cr, DUTY_PME) && !thisRankHasDuty(cr, DUTY_PP),
684 nthreads_pp, nthreads_pme);
685 auto cycle_sum(wallcycle_sum(cr, wcycle));
689 auto nbnxn_gpu_timings =
690 (nbv != nullptr && nbv->useGpu()) ? Nbnxm::gpu_get_timings(nbv->gpu_nbv) : nullptr;
691 gmx_wallclock_gpu_pme_t pme_gpu_timings = {};
693 if (pme_gpu_task_enabled(pme))
695 pme_gpu_get_timings(pme, &pme_gpu_timings);
697 wallcycle_print(fplog, mdlog, cr->nnodes, cr->npmenodes, nthreads_pp, nthreads_pme,
698 elapsed_time_over_all_ranks, wcycle, cycle_sum, nbnxn_gpu_timings,
701 if (EI_DYNAMICS(inputrec->eI))
703 delta_t = inputrec->delta_t;
708 print_perf(fplog, elapsed_time_over_all_threads_over_all_ranks, elapsed_time_over_all_ranks,
709 walltime_accounting_get_nsteps_done_since_reset(walltime_accounting),
710 delta_t, nbfs, mflop);
714 print_perf(stderr, elapsed_time_over_all_threads_over_all_ranks, elapsed_time_over_all_ranks,
715 walltime_accounting_get_nsteps_done_since_reset(walltime_accounting),
716 delta_t, nbfs, mflop);
721 int Mdrunner::mdrunner()
724 t_forcerec* fr = nullptr;
725 real ewaldcoeff_q = 0;
726 real ewaldcoeff_lj = 0;
727 int nChargePerturbed = -1, nTypePerturbed = 0;
728 gmx_wallcycle_t wcycle;
729 gmx_walltime_accounting_t walltime_accounting = nullptr;
730 MembedHolder membedHolder(filenames.size(), filenames.data());
731 gmx_hw_info_t* hwinfo = nullptr;
733 /* CAUTION: threads may be started later on in this function, so
734 cr doesn't reflect the final parallel state right now */
737 /* TODO: inputrec should tell us whether we use an algorithm, not a file option */
738 const bool doEssentialDynamics = opt2bSet("-ei", filenames.size(), filenames.data());
739 const bool doRerun = mdrunOptions.rerun;
741 // Handle task-assignment related user options.
742 EmulateGpuNonbonded emulateGpuNonbonded =
743 (getenv("GMX_EMULATE_GPU") != nullptr ? EmulateGpuNonbonded::Yes : EmulateGpuNonbonded::No);
745 std::vector<int> userGpuTaskAssignment;
748 userGpuTaskAssignment = parseUserTaskAssignmentString(hw_opt.userGpuTaskAssignment);
750 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
751 auto nonbondedTarget = findTaskTarget(nbpu_opt);
752 auto pmeTarget = findTaskTarget(pme_opt);
753 auto pmeFftTarget = findTaskTarget(pme_fft_opt);
754 auto bondedTarget = findTaskTarget(bonded_opt);
755 auto updateTarget = findTaskTarget(update_opt);
757 FILE* fplog = nullptr;
758 // If we are appending, we don't write log output because we need
759 // to check that the old log file matches what the checkpoint file
760 // expects. Otherwise, we should start to write log output now if
761 // there is a file ready for it.
762 if (logFileHandle != nullptr && startingBehavior != StartingBehavior::RestartWithAppending)
764 fplog = gmx_fio_getfp(logFileHandle);
766 const bool isSimulationMasterRank = findIsSimulationMasterRank(ms, simulationCommunicator);
767 gmx::LoggerOwner logOwner(buildLogger(fplog, isSimulationMasterRank));
768 gmx::MDLogger mdlog(logOwner.logger());
770 // TODO The thread-MPI master rank makes a working
771 // PhysicalNodeCommunicator here, but it gets rebuilt by all ranks
772 // after the threads have been launched. This works because no use
773 // is made of that communicator until after the execution paths
774 // have rejoined. But it is likely that we can improve the way
775 // this is expressed, e.g. by expressly running detection only the
776 // master rank for thread-MPI, rather than relying on the mutex
777 // and reference count.
778 PhysicalNodeCommunicator physicalNodeComm(libraryWorldCommunicator, gmx_physicalnode_id_hash());
779 hwinfo = gmx_detect_hardware(mdlog, physicalNodeComm);
781 gmx_print_detected_hardware(fplog, isSimulationMasterRank && isMasterSim(ms), mdlog, hwinfo);
783 std::vector<int> gpuIdsToUse = makeGpuIdsToUse(hwinfo->deviceInfoList, hw_opt.gpuIdsAvailable);
785 // Print citation requests after all software/hardware printing
786 pleaseCiteGromacs(fplog);
788 // Note: legacy program logic relies on checking whether these pointers are assigned.
789 // Objects may or may not be allocated later.
790 std::unique_ptr<t_inputrec> inputrec;
791 std::unique_ptr<t_state> globalState;
793 auto partialDeserializedTpr = std::make_unique<PartialDeserializedTprFile>();
795 if (isSimulationMasterRank)
797 // Allocate objects to be initialized by later function calls.
798 /* Only the master rank has the global state */
799 globalState = std::make_unique<t_state>();
800 inputrec = std::make_unique<t_inputrec>();
802 /* Read (nearly) all data required for the simulation
803 * and keep the partly serialized tpr contents to send to other ranks later
805 applyGlobalSimulationState(*inputHolder_.get(), partialDeserializedTpr.get(),
806 globalState.get(), inputrec.get(), &mtop);
809 /* Check and update the hardware options for internal consistency */
810 checkAndUpdateHardwareOptions(mdlog, &hw_opt, isSimulationMasterRank, domdecOptions.numPmeRanks,
813 if (GMX_THREAD_MPI && isSimulationMasterRank)
815 bool useGpuForNonbonded = false;
816 bool useGpuForPme = false;
819 GMX_RELEASE_ASSERT(inputrec != nullptr, "Keep the compiler happy");
821 // If the user specified the number of ranks, then we must
822 // respect that, but in default mode, we need to allow for
823 // the number of GPUs to choose the number of ranks.
824 auto canUseGpuForNonbonded = buildSupportsNonbondedOnGpu(nullptr);
825 useGpuForNonbonded = decideWhetherToUseGpusForNonbondedWithThreadMpi(
826 nonbondedTarget, gpuIdsToUse, userGpuTaskAssignment, emulateGpuNonbonded,
827 canUseGpuForNonbonded,
828 gpuAccelerationOfNonbondedIsUseful(mdlog, *inputrec, GMX_THREAD_MPI),
829 hw_opt.nthreads_tmpi);
830 useGpuForPme = decideWhetherToUseGpusForPmeWithThreadMpi(
831 useGpuForNonbonded, pmeTarget, gpuIdsToUse, userGpuTaskAssignment, *hwinfo,
832 *inputrec, hw_opt.nthreads_tmpi, domdecOptions.numPmeRanks);
834 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
836 /* Determine how many thread-MPI ranks to start.
838 * TODO Over-writing the user-supplied value here does
839 * prevent any possible subsequent checks from working
841 hw_opt.nthreads_tmpi =
842 get_nthreads_mpi(hwinfo, &hw_opt, gpuIdsToUse, useGpuForNonbonded, useGpuForPme,
843 inputrec.get(), &mtop, mdlog, membedHolder.doMembed());
845 // Now start the threads for thread MPI.
846 spawnThreads(hw_opt.nthreads_tmpi);
847 // The spawned threads enter mdrunner() and execution of
848 // master and spawned threads joins at the end of this block.
850 PhysicalNodeCommunicator(libraryWorldCommunicator, gmx_physicalnode_id_hash());
853 GMX_RELEASE_ASSERT(ms || simulationCommunicator != MPI_COMM_NULL,
854 "Must have valid communicator unless running a multi-simulation");
855 CommrecHandle crHandle = init_commrec(simulationCommunicator);
856 t_commrec* cr = crHandle.get();
857 GMX_RELEASE_ASSERT(cr != nullptr, "Must have valid commrec");
861 /* now broadcast everything to the non-master nodes/threads: */
862 if (!isSimulationMasterRank)
864 // Until now, only the master rank has a non-null pointer.
865 // On non-master ranks, allocate the object that will receive data in the following call.
866 inputrec = std::make_unique<t_inputrec>();
868 init_parallel(cr->mpiDefaultCommunicator, MASTER(cr), inputrec.get(), &mtop,
869 partialDeserializedTpr.get());
871 GMX_RELEASE_ASSERT(inputrec != nullptr, "All ranks should have a valid inputrec now");
872 partialDeserializedTpr.reset(nullptr);
874 // Now the number of ranks is known to all ranks, and each knows
875 // the inputrec read by the master rank. The ranks can now all run
876 // the task-deciding functions and will agree on the result
877 // without needing to communicate.
878 const bool useDomainDecomposition = (PAR(cr) && !(EI_TPI(inputrec->eI) || inputrec->eI == eiNM));
880 // Note that these variables describe only their own node.
882 // Note that when bonded interactions run on a GPU they always run
883 // alongside a nonbonded task, so do not influence task assignment
884 // even though they affect the force calculation workload.
885 bool useGpuForNonbonded = false;
886 bool useGpuForPme = false;
887 bool useGpuForBonded = false;
888 bool useGpuForUpdate = false;
889 bool gpusWereDetected = hwinfo->ngpu_compatible_tot > 0;
892 // It's possible that there are different numbers of GPUs on
893 // different nodes, which is the user's responsibility to
894 // handle. If unsuitable, we will notice that during task
896 auto canUseGpuForNonbonded = buildSupportsNonbondedOnGpu(nullptr);
897 useGpuForNonbonded = decideWhetherToUseGpusForNonbonded(
898 nonbondedTarget, userGpuTaskAssignment, emulateGpuNonbonded, canUseGpuForNonbonded,
899 gpuAccelerationOfNonbondedIsUseful(mdlog, *inputrec, !GMX_THREAD_MPI), gpusWereDetected);
900 useGpuForPme = decideWhetherToUseGpusForPme(
901 useGpuForNonbonded, pmeTarget, userGpuTaskAssignment, *hwinfo, *inputrec,
902 cr->sizeOfDefaultCommunicator, domdecOptions.numPmeRanks, gpusWereDetected);
903 auto canUseGpuForBonded = buildSupportsGpuBondeds(nullptr)
904 && inputSupportsGpuBondeds(*inputrec, mtop, nullptr);
905 useGpuForBonded = decideWhetherToUseGpusForBonded(
906 useGpuForNonbonded, useGpuForPme, bondedTarget, canUseGpuForBonded,
907 EVDW_PME(inputrec->vdwtype), EEL_PME_EWALD(inputrec->coulombtype),
908 domdecOptions.numPmeRanks, gpusWereDetected);
910 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
912 const PmeRunMode pmeRunMode = determinePmeRunMode(useGpuForPme, pmeFftTarget, *inputrec);
914 // Initialize development feature flags that enabled by environment variable
915 // and report those features that are enabled.
916 const DevelopmentFeatureFlags devFlags =
917 manageDevelopmentFeatures(mdlog, useGpuForNonbonded, pmeRunMode);
919 const bool useModularSimulator =
920 checkUseModularSimulator(false, inputrec.get(), doRerun, mtop, ms, replExParams,
921 nullptr, doEssentialDynamics, membedHolder.doMembed());
924 // TODO: hide restraint implementation details from Mdrunner.
925 // There is nothing unique about restraints at this point as far as the
926 // Mdrunner is concerned. The Mdrunner should just be getting a sequence of
927 // factory functions from the SimulationContext on which to call mdModules_->add().
928 // TODO: capture all restraints into a single RestraintModule, passed to the runner builder.
929 for (auto&& restraint : restraintManager_->getRestraints())
931 auto module = RestraintMDModule::create(restraint, restraint->sites());
932 mdModules_->add(std::move(module));
935 // TODO: Error handling
936 mdModules_->assignOptionsToModules(*inputrec->params, nullptr);
937 // now that the MdModules know their options, they know which callbacks to sign up to
938 mdModules_->subscribeToSimulationSetupNotifications();
939 const auto& mdModulesNotifier = mdModules_->notifier().simulationSetupNotifications_;
941 if (inputrec->internalParameters != nullptr)
943 mdModulesNotifier.notify(*inputrec->internalParameters);
946 if (fplog != nullptr)
948 pr_inputrec(fplog, 0, "Input Parameters", inputrec.get(), FALSE);
949 fprintf(fplog, "\n");
954 /* In rerun, set velocities to zero if present */
955 if (doRerun && ((globalState->flags & (1 << estV)) != 0))
957 // rerun does not use velocities
961 "Rerun trajectory contains velocities. Rerun does only evaluate "
962 "potential energy and forces. The velocities will be ignored.");
963 for (int i = 0; i < globalState->natoms; i++)
965 clear_rvec(globalState->v[i]);
967 globalState->flags &= ~(1 << estV);
970 /* now make sure the state is initialized and propagated */
971 set_state_entries(globalState.get(), inputrec.get(), useModularSimulator);
974 /* NM and TPI parallelize over force/energy calculations, not atoms,
975 * so we need to initialize and broadcast the global state.
977 if (inputrec->eI == eiNM || inputrec->eI == eiTPI)
981 globalState = std::make_unique<t_state>();
983 broadcastStateWithoutDynamics(cr->mpiDefaultCommunicator, DOMAINDECOMP(cr), PAR(cr),
987 /* A parallel command line option consistency check that we can
988 only do after any threads have started. */
990 && (domdecOptions.numCells[XX] > 1 || domdecOptions.numCells[YY] > 1
991 || domdecOptions.numCells[ZZ] > 1 || domdecOptions.numPmeRanks > 0))
994 "The -dd or -npme option request a parallel simulation, "
996 "but %s was compiled without threads or MPI enabled",
997 output_env_get_program_display_name(oenv));
999 "but the number of MPI-threads (option -ntmpi) is not set or is 1");
1001 "but %s was not started through mpirun/mpiexec or only one rank was requested "
1002 "through mpirun/mpiexec",
1003 output_env_get_program_display_name(oenv));
1007 if (doRerun && (EI_ENERGY_MINIMIZATION(inputrec->eI) || eiNM == inputrec->eI))
1010 "The .mdp file specified an energy mininization or normal mode algorithm, and "
1011 "these are not compatible with mdrun -rerun");
1014 if (!(EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype)))
1016 if (domdecOptions.numPmeRanks > 0)
1018 gmx_fatal_collective(FARGS, cr->mpiDefaultCommunicator, MASTER(cr),
1019 "PME-only ranks are requested, but the system does not use PME "
1020 "for electrostatics or LJ");
1023 domdecOptions.numPmeRanks = 0;
1026 if (useGpuForNonbonded && domdecOptions.numPmeRanks < 0)
1028 /* With NB GPUs we don't automatically use PME-only CPU ranks. PME ranks can
1029 * improve performance with many threads per GPU, since our OpenMP
1030 * scaling is bad, but it's difficult to automate the setup.
1032 domdecOptions.numPmeRanks = 0;
1036 if (domdecOptions.numPmeRanks < 0)
1038 domdecOptions.numPmeRanks = 0;
1039 // TODO possibly print a note that one can opt-in for a separate PME GPU rank?
1043 GMX_RELEASE_ASSERT(domdecOptions.numPmeRanks <= 1,
1044 "PME GPU decomposition is not supported");
1051 fcRegisterSteps(inputrec->nsteps, inputrec->init_step);
1055 /* NMR restraints must be initialized before load_checkpoint,
1056 * since with time averaging the history is added to t_state.
1057 * For proper consistency check we therefore need to extend
1059 * So the PME-only nodes (if present) will also initialize
1060 * the distance restraints.
1063 /* This needs to be called before read_checkpoint to extend the state */
1064 t_disresdata* disresdata;
1065 snew(disresdata, 1);
1066 init_disres(fplog, &mtop, inputrec.get(), DisResRunMode::MDRun,
1067 MASTER(cr) ? DDRole::Master : DDRole::Agent,
1068 PAR(cr) ? NumRanks::Multiple : NumRanks::Single, cr->mpi_comm_mysim, ms, disresdata,
1069 globalState.get(), replExParams.exchangeInterval > 0);
1071 t_oriresdata* oriresdata;
1072 snew(oriresdata, 1);
1073 init_orires(fplog, &mtop, inputrec.get(), cr, ms, globalState.get(), oriresdata);
1075 auto deform = prepareBoxDeformation(
1076 globalState != nullptr ? globalState->box : box, MASTER(cr) ? DDRole::Master : DDRole::Agent,
1077 PAR(cr) ? NumRanks::Multiple : NumRanks::Single, cr->mpi_comm_mygroup, *inputrec);
1079 ObservablesHistory observablesHistory = {};
1081 auto modularSimulatorCheckpointData = std::make_unique<ReadCheckpointDataHolder>();
1082 if (startingBehavior != StartingBehavior::NewSimulation)
1084 /* Check if checkpoint file exists before doing continuation.
1085 * This way we can use identical input options for the first and subsequent runs...
1087 if (mdrunOptions.numStepsCommandline > -2)
1089 /* Temporarily set the number of steps to unlimited to avoid
1090 * triggering the nsteps check in load_checkpoint().
1091 * This hack will go away soon when the -nsteps option is removed.
1093 inputrec->nsteps = -1;
1096 // Finish applying initial simulation state information from external sources on all ranks.
1097 // Reconcile checkpoint file data with Mdrunner state established up to this point.
1098 applyLocalState(*inputHolder_.get(), logFileHandle, cr, domdecOptions.numCells,
1099 inputrec.get(), globalState.get(), &observablesHistory,
1100 mdrunOptions.reproducible, mdModules_->notifier(),
1101 modularSimulatorCheckpointData.get(), useModularSimulator);
1102 // TODO: (#3652) Synchronize filesystem state, SimulationInput contents, and program
1104 // on all code paths.
1105 // Write checkpoint or provide hook to update SimulationInput.
1106 // If there was a checkpoint file, SimulationInput contains more information
1107 // than if there wasn't. At this point, we have synchronized the in-memory
1108 // state with the filesystem state only for restarted simulations. We should
1109 // be calling applyLocalState unconditionally and expect that the completeness
1110 // of SimulationInput is not dependent on its creation method.
1112 if (startingBehavior == StartingBehavior::RestartWithAppending && logFileHandle)
1114 // Now we can start normal logging to the truncated log file.
1115 fplog = gmx_fio_getfp(logFileHandle);
1116 prepareLogAppending(fplog);
1117 logOwner = buildLogger(fplog, MASTER(cr));
1118 mdlog = logOwner.logger();
1122 if (mdrunOptions.numStepsCommandline > -2)
1127 "The -nsteps functionality is deprecated, and may be removed in a future "
1129 "Consider using gmx convert-tpr -nsteps or changing the appropriate .mdp "
1132 /* override nsteps with value set on the commandline */
1133 override_nsteps_cmdline(mdlog, mdrunOptions.numStepsCommandline, inputrec.get());
1135 if (isSimulationMasterRank)
1137 copy_mat(globalState->box, box);
1142 gmx_bcast(sizeof(box), box, cr->mpiDefaultCommunicator);
1145 if (inputrec->cutoff_scheme != ecutsVERLET)
1148 "This group-scheme .tpr file can no longer be run by mdrun. Please update to the "
1149 "Verlet scheme, or use an earlier version of GROMACS if necessary.");
1151 /* Update rlist and nstlist. */
1152 /* Note: prepare_verlet_scheme is calling increaseNstlist(...), which (while attempting to
1153 * increase rlist) tries to check if the newly chosen value fits with the DD scheme. As this is
1154 * run before any DD scheme is set up, this check is never executed. See #3334 for more details.
1156 prepare_verlet_scheme(fplog, cr, inputrec.get(), nstlist_cmdline, &mtop, box,
1157 useGpuForNonbonded || (emulateGpuNonbonded == EmulateGpuNonbonded::Yes),
1160 // This builder is necessary while we have multi-part construction
1161 // of DD. Before DD is constructed, we use the existence of
1162 // the builder object to indicate that further construction of DD
1164 std::unique_ptr<DomainDecompositionBuilder> ddBuilder;
1165 if (useDomainDecomposition)
1167 ddBuilder = std::make_unique<DomainDecompositionBuilder>(
1168 mdlog, cr, domdecOptions, mdrunOptions, mtop, *inputrec, box,
1169 positionsFromStatePointer(globalState.get()));
1173 /* PME, if used, is done on all nodes with 1D decomposition */
1174 cr->nnodes = cr->sizeOfDefaultCommunicator;
1175 cr->sim_nodeid = cr->rankInDefaultCommunicator;
1176 cr->nodeid = cr->rankInDefaultCommunicator;
1178 cr->duty = (DUTY_PP | DUTY_PME);
1180 if (inputrec->pbcType == PbcType::Screw)
1182 gmx_fatal(FARGS, "pbc=screw is only implemented with domain decomposition");
1186 // Produce the task assignment for this rank - done after DD is constructed
1187 GpuTaskAssignments gpuTaskAssignments = GpuTaskAssignmentsBuilder::build(
1188 gpuIdsToUse, userGpuTaskAssignment, *hwinfo, simulationCommunicator, physicalNodeComm,
1189 nonbondedTarget, pmeTarget, bondedTarget, updateTarget, useGpuForNonbonded,
1190 useGpuForPme, thisRankHasDuty(cr, DUTY_PP),
1191 // TODO cr->duty & DUTY_PME should imply that a PME
1192 // algorithm is active, but currently does not.
1193 EEL_PME(inputrec->coulombtype) && thisRankHasDuty(cr, DUTY_PME));
1195 // Get the device handles for the modules, nullptr when no task is assigned.
1197 DeviceInformation* deviceInfo = gpuTaskAssignments.initDevice(&deviceId);
1199 // timing enabling - TODO put this in gpu_utils (even though generally this is just option handling?)
1200 bool useTiming = true;
1204 /* WARNING: CUDA timings are incorrect with multiple streams.
1205 * This is the main reason why they are disabled by default.
1207 // TODO: Consider turning on by default when we can detect nr of streams.
1208 useTiming = (getenv("GMX_ENABLE_GPU_TIMING") != nullptr);
1210 else if (GMX_GPU_OPENCL)
1212 useTiming = (getenv("GMX_DISABLE_GPU_TIMING") == nullptr);
1215 // TODO Currently this is always built, yet DD partition code
1216 // checks if it is built before using it. Probably it should
1217 // become an MDModule that is made only when another module
1218 // requires it (e.g. pull, CompEl, density fitting), so that we
1219 // don't update the local atom sets unilaterally every step.
1220 LocalAtomSetManager atomSets;
1223 // TODO Pass the GPU streams to ddBuilder to use in buffer
1224 // transfers (e.g. halo exchange)
1225 cr->dd = ddBuilder->build(&atomSets);
1226 // The builder's job is done, so destruct it
1227 ddBuilder.reset(nullptr);
1228 // Note that local state still does not exist yet.
1231 // The GPU update is decided here because we need to know whether the constraints or
1232 // SETTLEs can span accross the domain borders (i.e. whether or not update groups are
1233 // defined). This is only known after DD is initialized, hence decision on using GPU
1234 // update is done so late.
1237 const bool useUpdateGroups = cr->dd ? ddUsesUpdateGroups(*cr->dd) : false;
1239 useGpuForUpdate = decideWhetherToUseGpuForUpdate(
1240 useDomainDecomposition, useUpdateGroups, pmeRunMode, domdecOptions.numPmeRanks > 0,
1241 useGpuForNonbonded, updateTarget, gpusWereDetected, *inputrec, mtop,
1242 doEssentialDynamics, gmx_mtop_ftype_count(mtop, F_ORIRES) > 0,
1243 replExParams.exchangeInterval > 0, doRerun, devFlags, mdlog);
1245 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1247 const bool printHostName = (cr->nnodes > 1);
1248 gpuTaskAssignments.reportGpuUsage(mdlog, printHostName, useGpuForBonded, pmeRunMode, useGpuForUpdate);
1250 const bool disableNonbondedCalculation = (getenv("GMX_NO_NONBONDED") != nullptr);
1251 if (disableNonbondedCalculation)
1253 /* turn off non-bonded calculations */
1254 GMX_LOG(mdlog.warning)
1257 "Found environment variable GMX_NO_NONBONDED.\n"
1258 "Disabling nonbonded calculations.");
1261 MdrunScheduleWorkload runScheduleWork;
1263 bool useGpuDirectHalo = decideWhetherToUseGpuForHalo(
1264 devFlags, havePPDomainDecomposition(cr), useGpuForNonbonded, useModularSimulator,
1265 doRerun, EI_ENERGY_MINIMIZATION(inputrec->eI));
1267 // Also populates the simulation constant workload description.
1268 runScheduleWork.simulationWork = createSimulationWorkload(
1269 *inputrec, disableNonbondedCalculation, devFlags, useGpuForNonbonded, pmeRunMode,
1270 useGpuForBonded, useGpuForUpdate, useGpuDirectHalo);
1272 std::unique_ptr<DeviceStreamManager> deviceStreamManager = nullptr;
1274 if (deviceInfo != nullptr)
1276 if (DOMAINDECOMP(cr) && thisRankHasDuty(cr, DUTY_PP))
1278 dd_setup_dlb_resource_sharing(cr, deviceId);
1280 deviceStreamManager = std::make_unique<DeviceStreamManager>(
1281 *deviceInfo, havePPDomainDecomposition(cr), runScheduleWork.simulationWork, useTiming);
1284 // If the user chose a task assignment, give them some hints
1285 // where appropriate.
1286 if (!userGpuTaskAssignment.empty())
1288 gpuTaskAssignments.logPerformanceHints(mdlog, ssize(gpuIdsToUse));
1293 /* After possible communicator splitting in make_dd_communicators.
1294 * we can set up the intra/inter node communication.
1296 gmx_setup_nodecomm(fplog, cr);
1302 GMX_LOG(mdlog.warning)
1304 .appendTextFormatted(
1305 "This is simulation %d out of %d running as a composite GROMACS\n"
1306 "multi-simulation job. Setup for this simulation:\n",
1307 ms->simulationIndex_, ms->numSimulations_);
1309 GMX_LOG(mdlog.warning)
1310 .appendTextFormatted("Using %d MPI %s\n", cr->nnodes,
1312 cr->nnodes == 1 ? "thread" : "threads"
1314 cr->nnodes == 1 ? "process" : "processes"
1320 // If mdrun -pin auto honors any affinity setting that already
1321 // exists. If so, it is nice to provide feedback about whether
1322 // that existing affinity setting was from OpenMP or something
1323 // else, so we run this code both before and after we initialize
1324 // the OpenMP support.
1325 gmx_check_thread_affinity_set(mdlog, &hw_opt, hwinfo->nthreads_hw_avail, FALSE);
1326 /* Check and update the number of OpenMP threads requested */
1327 checkAndUpdateRequestedNumOpenmpThreads(&hw_opt, *hwinfo, cr, ms, physicalNodeComm.size_,
1328 pmeRunMode, mtop, *inputrec);
1330 gmx_omp_nthreads_init(mdlog, cr, hwinfo->nthreads_hw_avail, physicalNodeComm.size_,
1331 hw_opt.nthreads_omp, hw_opt.nthreads_omp_pme, !thisRankHasDuty(cr, DUTY_PP));
1333 // Enable FP exception detection, but not in
1334 // Release mode and not for compilers with known buggy FP
1335 // exception support (clang with any optimization) or suspected
1336 // buggy FP exception support (gcc 7.* with optimization).
1337 #if !defined NDEBUG \
1338 && !((defined __clang__ || (defined(__GNUC__) && !defined(__ICC) && __GNUC__ == 7)) \
1339 && defined __OPTIMIZE__)
1340 const bool bEnableFPE = true;
1342 const bool bEnableFPE = false;
1344 // FIXME - reconcile with gmx_feenableexcept() call from CommandLineModuleManager::run()
1347 gmx_feenableexcept();
1350 /* Now that we know the setup is consistent, check for efficiency */
1351 check_resource_division_efficiency(hwinfo, gpuTaskAssignments.thisRankHasAnyGpuTask(),
1352 mdrunOptions.ntompOptionIsSet, cr, mdlog);
1354 /* getting number of PP/PME threads on this MPI / tMPI rank.
1355 PME: env variable should be read only on one node to make sure it is
1356 identical everywhere;
1358 const int numThreadsOnThisRank = thisRankHasDuty(cr, DUTY_PP) ? gmx_omp_nthreads_get(emntNonbonded)
1359 : gmx_omp_nthreads_get(emntPME);
1360 checkHardwareOversubscription(numThreadsOnThisRank, cr->nodeid, *hwinfo->hardwareTopology,
1361 physicalNodeComm, mdlog);
1363 // Enable Peer access between GPUs where available
1364 // Only for DD, only master PP rank needs to perform setup, and only if thread MPI plus
1365 // any of the GPU communication features are active.
1366 if (DOMAINDECOMP(cr) && MASTER(cr) && thisRankHasDuty(cr, DUTY_PP) && GMX_THREAD_MPI
1367 && (runScheduleWork.simulationWork.useGpuHaloExchange
1368 || runScheduleWork.simulationWork.useGpuPmePpCommunication))
1370 setupGpuDevicePeerAccess(gpuIdsToUse, mdlog);
1373 if (hw_opt.threadAffinity != ThreadAffinity::Off)
1375 /* Before setting affinity, check whether the affinity has changed
1376 * - which indicates that probably the OpenMP library has changed it
1377 * since we first checked).
1379 gmx_check_thread_affinity_set(mdlog, &hw_opt, hwinfo->nthreads_hw_avail, TRUE);
1381 int numThreadsOnThisNode, intraNodeThreadOffset;
1382 analyzeThreadsOnThisNode(physicalNodeComm, numThreadsOnThisRank, &numThreadsOnThisNode,
1383 &intraNodeThreadOffset);
1385 /* Set the CPU affinity */
1386 gmx_set_thread_affinity(mdlog, cr, &hw_opt, *hwinfo->hardwareTopology, numThreadsOnThisRank,
1387 numThreadsOnThisNode, intraNodeThreadOffset, nullptr);
1390 if (mdrunOptions.timingOptions.resetStep > -1)
1395 "The -resetstep functionality is deprecated, and may be removed in a "
1398 wcycle = wallcycle_init(fplog, mdrunOptions.timingOptions.resetStep, cr);
1402 /* Master synchronizes its value of reset_counters with all nodes
1403 * including PME only nodes */
1404 int64_t reset_counters = wcycle_get_reset_counters(wcycle);
1405 gmx_bcast(sizeof(reset_counters), &reset_counters, cr->mpi_comm_mysim);
1406 wcycle_set_reset_counters(wcycle, reset_counters);
1409 // Membrane embedding must be initialized before we call init_forcerec()
1410 membedHolder.initializeMembed(fplog, filenames.size(), filenames.data(), &mtop, inputrec.get(),
1411 globalState.get(), cr, &mdrunOptions.checkpointOptions.period);
1413 const bool thisRankHasPmeGpuTask = gpuTaskAssignments.thisRankHasPmeGpuTask();
1414 std::unique_ptr<MDAtoms> mdAtoms;
1415 std::unique_ptr<VirtualSitesHandler> vsite;
1416 std::unique_ptr<GpuBonded> gpuBonded;
1419 if (thisRankHasDuty(cr, DUTY_PP))
1421 mdModulesNotifier.notify(*cr);
1422 mdModulesNotifier.notify(&atomSets);
1423 mdModulesNotifier.notify(inputrec->pbcType);
1424 mdModulesNotifier.notify(SimulationTimeStep{ inputrec->delta_t });
1425 /* Initiate forcerecord */
1426 fr = new t_forcerec;
1427 fr->forceProviders = mdModules_->initForceProviders();
1428 init_forcerec(fplog, mdlog, fr, inputrec.get(), &mtop, cr, box,
1429 opt2fn("-table", filenames.size(), filenames.data()),
1430 opt2fn("-tablep", filenames.size(), filenames.data()),
1431 opt2fns("-tableb", filenames.size(), filenames.data()), pforce);
1432 // Dirty hack, for fixing disres and orires should be made mdmodules
1433 fr->fcdata->disres = disresdata;
1434 fr->fcdata->orires = oriresdata;
1436 // Save a handle to device stream manager to use elsewhere in the code
1437 // TODO: Forcerec is not a correct place to store it.
1438 fr->deviceStreamManager = deviceStreamManager.get();
1440 if (runScheduleWork.simulationWork.useGpuPmePpCommunication && !thisRankHasDuty(cr, DUTY_PME))
1443 deviceStreamManager != nullptr,
1444 "GPU device stream manager should be valid in order to use PME-PP direct "
1447 deviceStreamManager->streamIsValid(DeviceStreamType::PmePpTransfer),
1448 "GPU PP-PME stream should be valid in order to use GPU PME-PP direct "
1450 fr->pmePpCommGpu = std::make_unique<gmx::PmePpCommGpu>(
1451 cr->mpi_comm_mysim, cr->dd->pme_nodeid, deviceStreamManager->context(),
1452 deviceStreamManager->stream(DeviceStreamType::PmePpTransfer));
1455 fr->nbv = Nbnxm::init_nb_verlet(mdlog, inputrec.get(), fr, cr, *hwinfo,
1456 runScheduleWork.simulationWork.useGpuNonbonded,
1457 deviceStreamManager.get(), &mtop, box, wcycle);
1458 // TODO: Move the logic below to a GPU bonded builder
1459 if (runScheduleWork.simulationWork.useGpuBonded)
1461 GMX_RELEASE_ASSERT(deviceStreamManager != nullptr,
1462 "GPU device stream manager should be valid in order to use GPU "
1463 "version of bonded forces.");
1464 gpuBonded = std::make_unique<GpuBonded>(
1465 mtop.ffparams, fr->ic->epsfac * fr->fudgeQQ, deviceStreamManager->context(),
1466 deviceStreamManager->bondedStream(havePPDomainDecomposition(cr)), wcycle);
1467 fr->gpuBonded = gpuBonded.get();
1470 /* Initialize the mdAtoms structure.
1471 * mdAtoms is not filled with atom data,
1472 * as this can not be done now with domain decomposition.
1474 mdAtoms = makeMDAtoms(fplog, mtop, *inputrec, thisRankHasPmeGpuTask);
1475 if (globalState && thisRankHasPmeGpuTask)
1477 // The pinning of coordinates in the global state object works, because we only use
1478 // PME on GPU without DD or on a separate PME rank, and because the local state pointer
1479 // points to the global state object without DD.
1480 // FIXME: MD and EM separately set up the local state - this should happen in the same
1481 // function, which should also perform the pinning.
1482 changePinningPolicy(&globalState->x, pme_get_pinning_policy());
1485 /* Initialize the virtual site communication */
1486 vsite = makeVirtualSitesHandler(mtop, cr, fr->pbcType);
1488 calc_shifts(box, fr->shift_vec);
1490 /* With periodic molecules the charge groups should be whole at start up
1491 * and the virtual sites should not be far from their proper positions.
1493 if (!inputrec->bContinuation && MASTER(cr)
1494 && !(inputrec->pbcType != PbcType::No && inputrec->bPeriodicMols))
1496 /* Make molecules whole at start of run */
1497 if (fr->pbcType != PbcType::No)
1499 do_pbc_first_mtop(fplog, inputrec->pbcType, box, &mtop, globalState->x.rvec_array());
1503 /* Correct initial vsite positions are required
1504 * for the initial distribution in the domain decomposition
1505 * and for the initial shell prediction.
1507 constructVirtualSitesGlobal(mtop, globalState->x);
1511 if (EEL_PME(fr->ic->eeltype) || EVDW_PME(fr->ic->vdwtype))
1513 ewaldcoeff_q = fr->ic->ewaldcoeff_q;
1514 ewaldcoeff_lj = fr->ic->ewaldcoeff_lj;
1519 /* This is a PME only node */
1521 GMX_ASSERT(globalState == nullptr,
1522 "We don't need the state on a PME only rank and expect it to be unitialized");
1524 ewaldcoeff_q = calc_ewaldcoeff_q(inputrec->rcoulomb, inputrec->ewald_rtol);
1525 ewaldcoeff_lj = calc_ewaldcoeff_lj(inputrec->rvdw, inputrec->ewald_rtol_lj);
1528 gmx_pme_t* sepPmeData = nullptr;
1529 // This reference hides the fact that PME data is owned by runner on PME-only ranks and by forcerec on other ranks
1530 GMX_ASSERT(thisRankHasDuty(cr, DUTY_PP) == (fr != nullptr),
1531 "Double-checking that only PME-only ranks have no forcerec");
1532 gmx_pme_t*& pmedata = fr ? fr->pmedata : sepPmeData;
1534 // TODO should live in ewald module once its testing is improved
1536 // Later, this program could contain kernels that might be later
1537 // re-used as auto-tuning progresses, or subsequent simulations
1539 PmeGpuProgramStorage pmeGpuProgram;
1540 if (thisRankHasPmeGpuTask)
1543 (deviceStreamManager != nullptr),
1544 "GPU device stream manager should be initialized in order to use GPU for PME.");
1545 GMX_RELEASE_ASSERT((deviceInfo != nullptr),
1546 "GPU device should be initialized in order to use GPU for PME.");
1547 pmeGpuProgram = buildPmeGpuProgram(deviceStreamManager->context());
1550 /* Initiate PME if necessary,
1551 * either on all nodes or on dedicated PME nodes only. */
1552 if (EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype))
1554 if (mdAtoms && mdAtoms->mdatoms())
1556 nChargePerturbed = mdAtoms->mdatoms()->nChargePerturbed;
1557 if (EVDW_PME(inputrec->vdwtype))
1559 nTypePerturbed = mdAtoms->mdatoms()->nTypePerturbed;
1562 if (cr->npmenodes > 0)
1564 /* The PME only nodes need to know nChargePerturbed(FEP on Q) and nTypePerturbed(FEP on LJ)*/
1565 gmx_bcast(sizeof(nChargePerturbed), &nChargePerturbed, cr->mpi_comm_mysim);
1566 gmx_bcast(sizeof(nTypePerturbed), &nTypePerturbed, cr->mpi_comm_mysim);
1569 if (thisRankHasDuty(cr, DUTY_PME))
1573 // TODO: This should be in the builder.
1574 GMX_RELEASE_ASSERT(!runScheduleWork.simulationWork.useGpuPme
1575 || (deviceStreamManager != nullptr),
1576 "Device stream manager should be valid in order to use GPU "
1579 !runScheduleWork.simulationWork.useGpuPme
1580 || deviceStreamManager->streamIsValid(DeviceStreamType::Pme),
1581 "GPU PME stream should be valid in order to use GPU version of PME.");
1583 const DeviceContext* deviceContext = runScheduleWork.simulationWork.useGpuPme
1584 ? &deviceStreamManager->context()
1586 const DeviceStream* pmeStream =
1587 runScheduleWork.simulationWork.useGpuPme
1588 ? &deviceStreamManager->stream(DeviceStreamType::Pme)
1591 pmedata = gmx_pme_init(cr, getNumPmeDomains(cr->dd), inputrec.get(),
1592 nChargePerturbed != 0, nTypePerturbed != 0,
1593 mdrunOptions.reproducible, ewaldcoeff_q, ewaldcoeff_lj,
1594 gmx_omp_nthreads_get(emntPME), pmeRunMode, nullptr,
1595 deviceContext, pmeStream, pmeGpuProgram.get(), mdlog);
1597 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1602 if (EI_DYNAMICS(inputrec->eI))
1604 /* Turn on signal handling on all nodes */
1606 * (A user signal from the PME nodes (if any)
1607 * is communicated to the PP nodes.
1609 signal_handler_install();
1612 pull_t* pull_work = nullptr;
1613 if (thisRankHasDuty(cr, DUTY_PP))
1615 /* Assumes uniform use of the number of OpenMP threads */
1616 walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(emntDefault));
1618 if (inputrec->bPull)
1620 /* Initialize pull code */
1621 pull_work = init_pull(fplog, inputrec->pull, inputrec.get(), &mtop, cr, &atomSets,
1622 inputrec->fepvals->init_lambda);
1623 if (inputrec->pull->bXOutAverage || inputrec->pull->bFOutAverage)
1625 initPullHistory(pull_work, &observablesHistory);
1627 if (EI_DYNAMICS(inputrec->eI) && MASTER(cr))
1629 init_pull_output_files(pull_work, filenames.size(), filenames.data(), oenv, startingBehavior);
1633 std::unique_ptr<EnforcedRotation> enforcedRotation;
1636 /* Initialize enforced rotation code */
1637 enforcedRotation = init_rot(fplog, inputrec.get(), filenames.size(), filenames.data(),
1638 cr, &atomSets, globalState.get(), &mtop, oenv, mdrunOptions,
1642 t_swap* swap = nullptr;
1643 if (inputrec->eSwapCoords != eswapNO)
1645 /* Initialize ion swapping code */
1646 swap = init_swapcoords(fplog, inputrec.get(),
1647 opt2fn_master("-swap", filenames.size(), filenames.data(), cr),
1648 &mtop, globalState.get(), &observablesHistory, cr, &atomSets,
1649 oenv, mdrunOptions, startingBehavior);
1652 /* Let makeConstraints know whether we have essential dynamics constraints. */
1653 auto constr = makeConstraints(mtop, *inputrec, pull_work, doEssentialDynamics, fplog, cr,
1654 ms, &nrnb, wcycle, fr->bMolPBC);
1656 /* Energy terms and groups */
1657 gmx_enerdata_t enerd(mtop.groups.groups[SimulationAtomGroupType::EnergyOutput].size(),
1658 inputrec->fepvals->n_lambda);
1660 /* Kinetic energy data */
1661 gmx_ekindata_t ekind;
1662 init_ekindata(fplog, &mtop, &(inputrec->opts), &ekind, inputrec->cos_accel);
1664 /* Set up interactive MD (IMD) */
1666 makeImdSession(inputrec.get(), cr, wcycle, &enerd, ms, &mtop, mdlog,
1667 MASTER(cr) ? globalState->x.rvec_array() : nullptr, filenames.size(),
1668 filenames.data(), oenv, mdrunOptions.imdOptions, startingBehavior);
1670 if (DOMAINDECOMP(cr))
1672 GMX_RELEASE_ASSERT(fr, "fr was NULL while cr->duty was DUTY_PP");
1673 /* This call is not included in init_domain_decomposition mainly
1674 * because fr->cginfo_mb is set later.
1676 dd_init_bondeds(fplog, cr->dd, mtop, vsite.get(), inputrec.get(),
1677 domdecOptions.checkBondedInteractions, fr->cginfo_mb);
1680 if (runScheduleWork.simulationWork.useGpuBufferOps)
1682 fr->gpuForceReduction[gmx::AtomLocality::Local] = std::make_unique<gmx::GpuForceReduction>(
1683 deviceStreamManager->context(),
1684 deviceStreamManager->stream(gmx::DeviceStreamType::NonBondedLocal));
1685 fr->gpuForceReduction[gmx::AtomLocality::NonLocal] = std::make_unique<gmx::GpuForceReduction>(
1686 deviceStreamManager->context(),
1687 deviceStreamManager->stream(gmx::DeviceStreamType::NonBondedNonLocal));
1690 std::unique_ptr<gmx::StatePropagatorDataGpu> stateGpu;
1691 if (gpusWereDetected
1692 && ((runScheduleWork.simulationWork.useGpuPme && thisRankHasDuty(cr, DUTY_PME))
1693 || runScheduleWork.simulationWork.useGpuBufferOps))
1695 GpuApiCallBehavior transferKind = (inputrec->eI == eiMD && !doRerun && !useModularSimulator)
1696 ? GpuApiCallBehavior::Async
1697 : GpuApiCallBehavior::Sync;
1698 GMX_RELEASE_ASSERT(deviceStreamManager != nullptr,
1699 "GPU device stream manager should be initialized to use GPU.");
1700 stateGpu = std::make_unique<gmx::StatePropagatorDataGpu>(
1701 *deviceStreamManager, transferKind, pme_gpu_get_block_size(fr->pmedata), wcycle);
1702 fr->stateGpu = stateGpu.get();
1705 GMX_ASSERT(stopHandlerBuilder_, "Runner must provide StopHandlerBuilder to simulator.");
1706 SimulatorBuilder simulatorBuilder;
1708 simulatorBuilder.add(SimulatorStateData(globalState.get(), &observablesHistory, &enerd, &ekind));
1709 simulatorBuilder.add(std::move(membedHolder));
1710 simulatorBuilder.add(std::move(stopHandlerBuilder_));
1711 simulatorBuilder.add(SimulatorConfig(mdrunOptions, startingBehavior, &runScheduleWork));
1714 simulatorBuilder.add(SimulatorEnv(fplog, cr, ms, mdlog, oenv));
1715 simulatorBuilder.add(Profiling(&nrnb, walltime_accounting, wcycle));
1716 simulatorBuilder.add(ConstraintsParam(
1717 constr.get(), enforcedRotation ? enforcedRotation->getLegacyEnfrot() : nullptr,
1719 // TODO: Separate `fr` to a separate add, and make the `build` handle the coupling sensibly.
1720 simulatorBuilder.add(LegacyInput(static_cast<int>(filenames.size()), filenames.data(),
1721 inputrec.get(), fr));
1722 simulatorBuilder.add(ReplicaExchangeParameters(replExParams));
1723 simulatorBuilder.add(InteractiveMD(imdSession.get()));
1724 simulatorBuilder.add(SimulatorModules(mdModules_->outputProvider(), mdModules_->notifier()));
1725 simulatorBuilder.add(CenterOfMassPulling(pull_work));
1726 // Todo move to an MDModule
1727 simulatorBuilder.add(IonSwapping(swap));
1728 simulatorBuilder.add(TopologyData(&mtop, mdAtoms.get()));
1729 simulatorBuilder.add(BoxDeformationHandle(deform.get()));
1730 simulatorBuilder.add(std::move(modularSimulatorCheckpointData));
1732 // build and run simulator object based on user-input
1733 auto simulator = simulatorBuilder.build(useModularSimulator);
1736 if (fr->pmePpCommGpu)
1738 // destroy object since it is no longer required. (This needs to be done while the GPU context still exists.)
1739 fr->pmePpCommGpu.reset();
1742 if (inputrec->bPull)
1744 finish_pull(pull_work);
1746 finish_swapcoords(swap);
1750 GMX_RELEASE_ASSERT(pmedata, "pmedata was NULL while cr->duty was not DUTY_PP");
1752 walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(emntPME));
1753 gmx_pmeonly(pmedata, cr, &nrnb, wcycle, walltime_accounting, inputrec.get(), pmeRunMode,
1754 deviceStreamManager.get());
1757 wallcycle_stop(wcycle, ewcRUN);
1759 /* Finish up, write some stuff
1760 * if rerunMD, don't write last frame again
1762 finish_run(fplog, mdlog, cr, inputrec.get(), &nrnb, wcycle, walltime_accounting,
1763 fr ? fr->nbv.get() : nullptr, pmedata, EI_DYNAMICS(inputrec->eI) && !isMultiSim(ms));
1765 // clean up cycle counter
1766 wallcycle_destroy(wcycle);
1768 deviceStreamManager.reset(nullptr);
1772 gmx_pme_destroy(pmedata);
1776 // FIXME: this is only here to manually unpin mdAtoms->chargeA_ and state->x,
1777 // before we destroy the GPU context(s)
1778 // Pinned buffers are associated with contexts in CUDA.
1779 // As soon as we destroy GPU contexts after mdrunner() exits, these lines should go.
1780 mdAtoms.reset(nullptr);
1781 globalState.reset(nullptr);
1782 mdModules_.reset(nullptr); // destruct force providers here as they might also use the GPU
1783 gpuBonded.reset(nullptr);
1784 /* Free pinned buffers in *fr */
1787 // TODO convert to C++ so we can get rid of these frees
1791 if (!hwinfo->deviceInfoList.empty())
1793 /* stop the GPU profiler (only CUDA) */
1797 /* With tMPI we need to wait for all ranks to finish deallocation before
1798 * destroying the CUDA context as some tMPI ranks may be sharing
1801 * This is not a concern in OpenCL where we use one context per rank.
1803 * Note: it is safe to not call the barrier on the ranks which do not use GPU,
1804 * but it is easier and more futureproof to call it on the whole node.
1806 * Note that this function needs to be called even if GPUs are not used
1807 * in this run because the PME ranks have no knowledge of whether GPUs
1808 * are used or not, but all ranks need to enter the barrier below.
1809 * \todo Remove this physical node barrier after making sure
1810 * that it's not needed anymore (with a shared GPU run).
1814 physicalNodeComm.barrier();
1816 releaseDevice(deviceInfo);
1818 /* Does what it says */
1819 print_date_and_time(fplog, cr->nodeid, "Finished mdrun", gmx_gettime());
1820 walltime_accounting_destroy(walltime_accounting);
1822 // Ensure log file content is written
1825 gmx_fio_flush(logFileHandle);
1828 /* Reset FPEs (important for unit tests) by disabling them. Assumes no
1829 * exceptions were enabled before function was called. */
1832 gmx_fedisableexcept();
1835 auto rc = static_cast<int>(gmx_get_stop_condition());
1838 /* we need to join all threads. The sub-threads join when they
1839 exit this function, but the master thread needs to be told to
1849 Mdrunner::~Mdrunner()
1851 // Clean up of the Manager.
1852 // This will end up getting called on every thread-MPI rank, which is unnecessary,
1853 // but okay as long as threads synchronize some time before adding or accessing
1854 // a new set of restraints.
1855 if (restraintManager_)
1857 restraintManager_->clear();
1858 GMX_ASSERT(restraintManager_->countRestraints() == 0,
1859 "restraints added during runner life time should be cleared at runner "
1864 void Mdrunner::addPotential(std::shared_ptr<gmx::IRestraintPotential> puller, const std::string& name)
1866 GMX_ASSERT(restraintManager_, "Mdrunner must have a restraint manager.");
1867 // Not sure if this should be logged through the md logger or something else,
1868 // but it is helpful to have some sort of INFO level message sent somewhere.
1869 // std::cout << "Registering restraint named " << name << std::endl;
1871 // When multiple restraints are used, it may be wasteful to register them separately.
1872 // Maybe instead register an entire Restraint Manager as a force provider.
1873 restraintManager_->addToSpec(std::move(puller), name);
1876 Mdrunner::Mdrunner(std::unique_ptr<MDModules> mdModules) : mdModules_(std::move(mdModules)) {}
1878 Mdrunner::Mdrunner(Mdrunner&&) noexcept = default;
1880 Mdrunner& Mdrunner::operator=(Mdrunner&& /*handle*/) noexcept = default;
1882 class Mdrunner::BuilderImplementation
1885 BuilderImplementation() = delete;
1886 BuilderImplementation(std::unique_ptr<MDModules> mdModules, compat::not_null<SimulationContext*> context);
1887 ~BuilderImplementation();
1889 BuilderImplementation& setExtraMdrunOptions(const MdrunOptions& options,
1890 real forceWarningThreshold,
1891 StartingBehavior startingBehavior);
1893 void addDomdec(const DomdecOptions& options);
1895 void addInput(SimulationInputHandle inputHolder);
1897 void addVerletList(int nstlist);
1899 void addReplicaExchange(const ReplicaExchangeParameters& params);
1901 void addNonBonded(const char* nbpu_opt);
1903 void addPME(const char* pme_opt_, const char* pme_fft_opt_);
1905 void addBondedTaskAssignment(const char* bonded_opt);
1907 void addUpdateTaskAssignment(const char* update_opt);
1909 void addHardwareOptions(const gmx_hw_opt_t& hardwareOptions);
1911 void addFilenames(ArrayRef<const t_filenm> filenames);
1913 void addOutputEnvironment(gmx_output_env_t* outputEnvironment);
1915 void addLogFile(t_fileio* logFileHandle);
1917 void addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder);
1922 // Default parameters copied from runner.h
1923 // \todo Clarify source(s) of default parameters.
1925 const char* nbpu_opt_ = nullptr;
1926 const char* pme_opt_ = nullptr;
1927 const char* pme_fft_opt_ = nullptr;
1928 const char* bonded_opt_ = nullptr;
1929 const char* update_opt_ = nullptr;
1931 MdrunOptions mdrunOptions_;
1933 DomdecOptions domdecOptions_;
1935 ReplicaExchangeParameters replicaExchangeParameters_;
1937 //! Command-line override for the duration of a neighbor list with the Verlet scheme.
1940 //! World communicator, used for hardware detection and task assignment
1941 MPI_Comm libraryWorldCommunicator_ = MPI_COMM_NULL;
1943 //! Multisim communicator handle.
1944 gmx_multisim_t* multiSimulation_;
1946 //! mdrun communicator
1947 MPI_Comm simulationCommunicator_ = MPI_COMM_NULL;
1949 //! Print a warning if any force is larger than this (in kJ/mol nm).
1950 real forceWarningThreshold_ = -1;
1952 //! Whether the simulation will start afresh, or restart with/without appending.
1953 StartingBehavior startingBehavior_ = StartingBehavior::NewSimulation;
1955 //! The modules that comprise the functionality of mdrun.
1956 std::unique_ptr<MDModules> mdModules_;
1958 //! \brief Parallelism information.
1959 gmx_hw_opt_t hardwareOptions_;
1961 //! filename options for simulation.
1962 ArrayRef<const t_filenm> filenames_;
1964 /*! \brief Handle to output environment.
1966 * \todo gmx_output_env_t needs lifetime management.
1968 gmx_output_env_t* outputEnvironment_ = nullptr;
1970 /*! \brief Non-owning handle to MD log file.
1972 * \todo Context should own output facilities for client.
1973 * \todo Improve log file handle management.
1975 * Code managing the FILE* relies on the ability to set it to
1976 * nullptr to check whether the filehandle is valid.
1978 t_fileio* logFileHandle_ = nullptr;
1981 * \brief Builder for simulation stop signal handler.
1983 std::unique_ptr<StopHandlerBuilder> stopHandlerBuilder_ = nullptr;
1986 * \brief Sources for initial simulation state.
1988 * See issue #3652 for near-term refinements to the SimulationInput interface.
1990 * See issue #3379 for broader discussion on API aspects of simulation inputs and outputs.
1992 SimulationInputHandle inputHolder_;
1995 Mdrunner::BuilderImplementation::BuilderImplementation(std::unique_ptr<MDModules> mdModules,
1996 compat::not_null<SimulationContext*> context) :
1997 mdModules_(std::move(mdModules))
1999 libraryWorldCommunicator_ = context->libraryWorldCommunicator_;
2000 simulationCommunicator_ = context->simulationCommunicator_;
2001 multiSimulation_ = context->multiSimulation_.get();
2004 Mdrunner::BuilderImplementation::~BuilderImplementation() = default;
2006 Mdrunner::BuilderImplementation&
2007 Mdrunner::BuilderImplementation::setExtraMdrunOptions(const MdrunOptions& options,
2008 const real forceWarningThreshold,
2009 const StartingBehavior startingBehavior)
2011 mdrunOptions_ = options;
2012 forceWarningThreshold_ = forceWarningThreshold;
2013 startingBehavior_ = startingBehavior;
2017 void Mdrunner::BuilderImplementation::addDomdec(const DomdecOptions& options)
2019 domdecOptions_ = options;
2022 void Mdrunner::BuilderImplementation::addVerletList(int nstlist)
2027 void Mdrunner::BuilderImplementation::addReplicaExchange(const ReplicaExchangeParameters& params)
2029 replicaExchangeParameters_ = params;
2032 Mdrunner Mdrunner::BuilderImplementation::build()
2034 auto newRunner = Mdrunner(std::move(mdModules_));
2036 newRunner.mdrunOptions = mdrunOptions_;
2037 newRunner.pforce = forceWarningThreshold_;
2038 newRunner.startingBehavior = startingBehavior_;
2039 newRunner.domdecOptions = domdecOptions_;
2041 // \todo determine an invariant to check or confirm that all gmx_hw_opt_t objects are valid
2042 newRunner.hw_opt = hardwareOptions_;
2044 // No invariant to check. This parameter exists to optionally override other behavior.
2045 newRunner.nstlist_cmdline = nstlist_;
2047 newRunner.replExParams = replicaExchangeParameters_;
2049 newRunner.filenames = filenames_;
2051 newRunner.libraryWorldCommunicator = libraryWorldCommunicator_;
2053 newRunner.simulationCommunicator = simulationCommunicator_;
2055 // nullptr is a valid value for the multisim handle
2056 newRunner.ms = multiSimulation_;
2060 newRunner.inputHolder_ = std::move(inputHolder_);
2064 GMX_THROW(gmx::APIError("MdrunnerBuilder::addInput() is required before build()."));
2067 // \todo Clarify ownership and lifetime management for gmx_output_env_t
2068 // \todo Update sanity checking when output environment has clearly specified invariants.
2069 // Initialization and default values for oenv are not well specified in the current version.
2070 if (outputEnvironment_)
2072 newRunner.oenv = outputEnvironment_;
2076 GMX_THROW(gmx::APIError(
2077 "MdrunnerBuilder::addOutputEnvironment() is required before build()"));
2080 newRunner.logFileHandle = logFileHandle_;
2084 newRunner.nbpu_opt = nbpu_opt_;
2088 GMX_THROW(gmx::APIError("MdrunnerBuilder::addNonBonded() is required before build()"));
2091 if (pme_opt_ && pme_fft_opt_)
2093 newRunner.pme_opt = pme_opt_;
2094 newRunner.pme_fft_opt = pme_fft_opt_;
2098 GMX_THROW(gmx::APIError("MdrunnerBuilder::addElectrostatics() is required before build()"));
2103 newRunner.bonded_opt = bonded_opt_;
2107 GMX_THROW(gmx::APIError(
2108 "MdrunnerBuilder::addBondedTaskAssignment() is required before build()"));
2113 newRunner.update_opt = update_opt_;
2117 GMX_THROW(gmx::APIError(
2118 "MdrunnerBuilder::addUpdateTaskAssignment() is required before build() "));
2122 newRunner.restraintManager_ = std::make_unique<gmx::RestraintManager>();
2124 if (stopHandlerBuilder_)
2126 newRunner.stopHandlerBuilder_ = std::move(stopHandlerBuilder_);
2130 newRunner.stopHandlerBuilder_ = std::make_unique<StopHandlerBuilder>();
2136 void Mdrunner::BuilderImplementation::addNonBonded(const char* nbpu_opt)
2138 nbpu_opt_ = nbpu_opt;
2141 void Mdrunner::BuilderImplementation::addPME(const char* pme_opt, const char* pme_fft_opt)
2144 pme_fft_opt_ = pme_fft_opt;
2147 void Mdrunner::BuilderImplementation::addBondedTaskAssignment(const char* bonded_opt)
2149 bonded_opt_ = bonded_opt;
2152 void Mdrunner::BuilderImplementation::addUpdateTaskAssignment(const char* update_opt)
2154 update_opt_ = update_opt;
2157 void Mdrunner::BuilderImplementation::addHardwareOptions(const gmx_hw_opt_t& hardwareOptions)
2159 hardwareOptions_ = hardwareOptions;
2162 void Mdrunner::BuilderImplementation::addFilenames(ArrayRef<const t_filenm> filenames)
2164 filenames_ = filenames;
2167 void Mdrunner::BuilderImplementation::addOutputEnvironment(gmx_output_env_t* outputEnvironment)
2169 outputEnvironment_ = outputEnvironment;
2172 void Mdrunner::BuilderImplementation::addLogFile(t_fileio* logFileHandle)
2174 logFileHandle_ = logFileHandle;
2177 void Mdrunner::BuilderImplementation::addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder)
2179 stopHandlerBuilder_ = std::move(builder);
2182 void Mdrunner::BuilderImplementation::addInput(SimulationInputHandle inputHolder)
2184 inputHolder_ = std::move(inputHolder);
2187 MdrunnerBuilder::MdrunnerBuilder(std::unique_ptr<MDModules> mdModules,
2188 compat::not_null<SimulationContext*> context) :
2189 impl_{ std::make_unique<Mdrunner::BuilderImplementation>(std::move(mdModules), context) }
2193 MdrunnerBuilder::~MdrunnerBuilder() = default;
2195 MdrunnerBuilder& MdrunnerBuilder::addSimulationMethod(const MdrunOptions& options,
2196 real forceWarningThreshold,
2197 const StartingBehavior startingBehavior)
2199 impl_->setExtraMdrunOptions(options, forceWarningThreshold, startingBehavior);
2203 MdrunnerBuilder& MdrunnerBuilder::addDomainDecomposition(const DomdecOptions& options)
2205 impl_->addDomdec(options);
2209 MdrunnerBuilder& MdrunnerBuilder::addNeighborList(int nstlist)
2211 impl_->addVerletList(nstlist);
2215 MdrunnerBuilder& MdrunnerBuilder::addReplicaExchange(const ReplicaExchangeParameters& params)
2217 impl_->addReplicaExchange(params);
2221 MdrunnerBuilder& MdrunnerBuilder::addNonBonded(const char* nbpu_opt)
2223 impl_->addNonBonded(nbpu_opt);
2227 MdrunnerBuilder& MdrunnerBuilder::addElectrostatics(const char* pme_opt, const char* pme_fft_opt)
2229 // The builder method may become more general in the future, but in this version,
2230 // parameters for PME electrostatics are both required and the only parameters
2232 if (pme_opt && pme_fft_opt)
2234 impl_->addPME(pme_opt, pme_fft_opt);
2239 gmx::InvalidInputError("addElectrostatics() arguments must be non-null pointers."));
2244 MdrunnerBuilder& MdrunnerBuilder::addBondedTaskAssignment(const char* bonded_opt)
2246 impl_->addBondedTaskAssignment(bonded_opt);
2250 MdrunnerBuilder& MdrunnerBuilder::addUpdateTaskAssignment(const char* update_opt)
2252 impl_->addUpdateTaskAssignment(update_opt);
2256 Mdrunner MdrunnerBuilder::build()
2258 return impl_->build();
2261 MdrunnerBuilder& MdrunnerBuilder::addHardwareOptions(const gmx_hw_opt_t& hardwareOptions)
2263 impl_->addHardwareOptions(hardwareOptions);
2267 MdrunnerBuilder& MdrunnerBuilder::addFilenames(ArrayRef<const t_filenm> filenames)
2269 impl_->addFilenames(filenames);
2273 MdrunnerBuilder& MdrunnerBuilder::addOutputEnvironment(gmx_output_env_t* outputEnvironment)
2275 impl_->addOutputEnvironment(outputEnvironment);
2279 MdrunnerBuilder& MdrunnerBuilder::addLogFile(t_fileio* logFileHandle)
2281 impl_->addLogFile(logFileHandle);
2285 MdrunnerBuilder& MdrunnerBuilder::addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder)
2287 impl_->addStopHandlerBuilder(std::move(builder));
2291 MdrunnerBuilder& MdrunnerBuilder::addInput(SimulationInputHandle input)
2293 impl_->addInput(std::move(input));
2297 MdrunnerBuilder::MdrunnerBuilder(MdrunnerBuilder&&) noexcept = default;
2299 MdrunnerBuilder& MdrunnerBuilder::operator=(MdrunnerBuilder&&) noexcept = default;