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) 2011,2012,2013,2014,2015,2016,2017,2018,2019, by the GROMACS development team, led by
7 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
8 * and including many others, as listed in the AUTHORS file in the
9 * top-level source directory and at http://www.gromacs.org.
11 * GROMACS is free software; you can redistribute it and/or
12 * modify it under the terms of the GNU Lesser General Public License
13 * as published by the Free Software Foundation; either version 2.1
14 * of the License, or (at your option) any later version.
16 * GROMACS is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
19 * Lesser General Public License for more details.
21 * You should have received a copy of the GNU Lesser General Public
22 * License along with GROMACS; if not, see
23 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
24 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
26 * If you want to redistribute modifications to GROMACS, please
27 * consider that scientific software is very special. Version
28 * control is crucial - bugs must be traceable. We will be happy to
29 * consider code for inclusion in the official distribution, but
30 * derived work must not be called official GROMACS. Details are found
31 * in the README & COPYING files - if they are missing, get the
32 * official version at http://www.gromacs.org.
34 * To help us fund GROMACS development, we humbly ask that you cite
35 * the research papers on the package. Check out http://www.gromacs.org.
39 * \brief 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/domdec.h"
61 #include "gromacs/domdec/domdec_struct.h"
62 #include "gromacs/domdec/localatomsetmanager.h"
63 #include "gromacs/domdec/partition.h"
64 #include "gromacs/ewald/ewald_utils.h"
65 #include "gromacs/ewald/pme.h"
66 #include "gromacs/ewald/pme_gpu_program.h"
67 #include "gromacs/fileio/checkpoint.h"
68 #include "gromacs/fileio/gmxfio.h"
69 #include "gromacs/fileio/oenv.h"
70 #include "gromacs/fileio/tpxio.h"
71 #include "gromacs/gmxlib/network.h"
72 #include "gromacs/gmxlib/nrnb.h"
73 #include "gromacs/gpu_utils/clfftinitializer.h"
74 #include "gromacs/gpu_utils/gpu_utils.h"
75 #include "gromacs/hardware/cpuinfo.h"
76 #include "gromacs/hardware/detecthardware.h"
77 #include "gromacs/hardware/printhardware.h"
78 #include "gromacs/imd/imd.h"
79 #include "gromacs/listed_forces/disre.h"
80 #include "gromacs/listed_forces/gpubonded.h"
81 #include "gromacs/listed_forces/orires.h"
82 #include "gromacs/math/functions.h"
83 #include "gromacs/math/utilities.h"
84 #include "gromacs/math/vec.h"
85 #include "gromacs/mdlib/boxdeformation.h"
86 #include "gromacs/mdlib/broadcaststructs.h"
87 #include "gromacs/mdlib/calc_verletbuf.h"
88 #include "gromacs/mdlib/dispersioncorrection.h"
89 #include "gromacs/mdlib/enerdata_utils.h"
90 #include "gromacs/mdlib/force.h"
91 #include "gromacs/mdlib/forcerec.h"
92 #include "gromacs/mdlib/gmx_omp_nthreads.h"
93 #include "gromacs/mdlib/makeconstraints.h"
94 #include "gromacs/mdlib/md_support.h"
95 #include "gromacs/mdlib/mdatoms.h"
96 #include "gromacs/mdlib/membed.h"
97 #include "gromacs/mdlib/ppforceworkload.h"
98 #include "gromacs/mdlib/qmmm.h"
99 #include "gromacs/mdlib/sighandler.h"
100 #include "gromacs/mdlib/stophandler.h"
101 #include "gromacs/mdrun/mdmodulenotification.h"
102 #include "gromacs/mdrun/mdmodules.h"
103 #include "gromacs/mdrun/simulationcontext.h"
104 #include "gromacs/mdrunutility/handlerestart.h"
105 #include "gromacs/mdrunutility/logging.h"
106 #include "gromacs/mdrunutility/multisim.h"
107 #include "gromacs/mdrunutility/printtime.h"
108 #include "gromacs/mdrunutility/threadaffinity.h"
109 #include "gromacs/mdtypes/commrec.h"
110 #include "gromacs/mdtypes/enerdata.h"
111 #include "gromacs/mdtypes/fcdata.h"
112 #include "gromacs/mdtypes/group.h"
113 #include "gromacs/mdtypes/inputrec.h"
114 #include "gromacs/mdtypes/md_enums.h"
115 #include "gromacs/mdtypes/mdrunoptions.h"
116 #include "gromacs/mdtypes/observableshistory.h"
117 #include "gromacs/mdtypes/state.h"
118 #include "gromacs/nbnxm/gpu_data_mgmt.h"
119 #include "gromacs/nbnxm/nbnxm.h"
120 #include "gromacs/nbnxm/pairlist_tuning.h"
121 #include "gromacs/pbcutil/pbc.h"
122 #include "gromacs/pulling/output.h"
123 #include "gromacs/pulling/pull.h"
124 #include "gromacs/pulling/pull_rotation.h"
125 #include "gromacs/restraint/manager.h"
126 #include "gromacs/restraint/restraintmdmodule.h"
127 #include "gromacs/restraint/restraintpotential.h"
128 #include "gromacs/swap/swapcoords.h"
129 #include "gromacs/taskassignment/decidegpuusage.h"
130 #include "gromacs/taskassignment/resourcedivision.h"
131 #include "gromacs/taskassignment/taskassignment.h"
132 #include "gromacs/taskassignment/usergpuids.h"
133 #include "gromacs/timing/gpu_timing.h"
134 #include "gromacs/timing/wallcycle.h"
135 #include "gromacs/timing/wallcyclereporting.h"
136 #include "gromacs/topology/mtop_util.h"
137 #include "gromacs/trajectory/trajectoryframe.h"
138 #include "gromacs/utility/basenetwork.h"
139 #include "gromacs/utility/cstringutil.h"
140 #include "gromacs/utility/exceptions.h"
141 #include "gromacs/utility/fatalerror.h"
142 #include "gromacs/utility/filestream.h"
143 #include "gromacs/utility/gmxassert.h"
144 #include "gromacs/utility/gmxmpi.h"
145 #include "gromacs/utility/keyvaluetree.h"
146 #include "gromacs/utility/logger.h"
147 #include "gromacs/utility/loggerbuilder.h"
148 #include "gromacs/utility/physicalnodecommunicator.h"
149 #include "gromacs/utility/pleasecite.h"
150 #include "gromacs/utility/programcontext.h"
151 #include "gromacs/utility/smalloc.h"
152 #include "gromacs/utility/stringutil.h"
154 #include "isimulator.h"
155 #include "replicaexchange.h"
156 #include "simulatorbuilder.h"
159 #include "corewrap.h"
165 /*! \brief Log if development feature flags are encountered
167 * The use of dev features indicated by environment variables is logged
168 * in order to ensure that runs with such featrues enabled can be identified
169 * from their log and standard output.
171 * \param[in] mdlog Logger object.
173 static void reportDevelopmentFeatures(const gmx::MDLogger &mdlog)
175 const bool enableGpuBufOps = (getenv("GMX_USE_GPU_BUFFER_OPS") != nullptr);
176 const bool useGpuUpdateConstrain = (getenv("GMX_UPDATE_CONSTRAIN_GPU") != nullptr);
180 GMX_LOG(mdlog.warning).asParagraph().appendTextFormatted(
181 "NOTE: This run uses the 'GPU buffer ops' feature, enabled by the GMX_USE_GPU_BUFFER_OPS environment variable.");
184 if (useGpuUpdateConstrain)
186 GMX_LOG(mdlog.warning).asParagraph().appendTextFormatted(
187 "NOTE: This run uses the 'GPU update/constraints' feature, enabled by the GMX_UPDATE_CONSTRAIN_GPU environment variable.");
191 /*! \brief Barrier for safe simultaneous thread access to mdrunner data
193 * Used to ensure that the master thread does not modify mdrunner during copy
194 * on the spawned threads. */
195 static void threadMpiMdrunnerAccessBarrier()
198 MPI_Barrier(MPI_COMM_WORLD);
202 Mdrunner Mdrunner::cloneOnSpawnedThread() const
204 auto newRunner = Mdrunner(std::make_unique<MDModules>());
206 // All runners in the same process share a restraint manager resource because it is
207 // part of the interface to the client code, which is associated only with the
208 // original thread. Handles to the same resources can be obtained by copy.
210 newRunner.restraintManager_ = std::make_unique<RestraintManager>(*restraintManager_);
213 // Copy original cr pointer before master thread can pass the thread barrier
214 newRunner.cr = reinitialize_commrec_for_this_thread(cr);
216 // Copy members of master runner.
217 // \todo Replace with builder when Simulation context and/or runner phases are better defined.
218 // Ref https://redmine.gromacs.org/issues/2587 and https://redmine.gromacs.org/issues/2375
219 newRunner.hw_opt = hw_opt;
220 newRunner.filenames = filenames;
222 newRunner.oenv = oenv;
223 newRunner.mdrunOptions = mdrunOptions;
224 newRunner.domdecOptions = domdecOptions;
225 newRunner.nbpu_opt = nbpu_opt;
226 newRunner.pme_opt = pme_opt;
227 newRunner.pme_fft_opt = pme_fft_opt;
228 newRunner.bonded_opt = bonded_opt;
229 newRunner.nstlist_cmdline = nstlist_cmdline;
230 newRunner.replExParams = replExParams;
231 newRunner.pforce = pforce;
233 newRunner.startingBehavior = startingBehavior;
234 newRunner.stopHandlerBuilder_ = std::make_unique<StopHandlerBuilder>(*stopHandlerBuilder_);
236 threadMpiMdrunnerAccessBarrier();
238 GMX_RELEASE_ASSERT(!MASTER(newRunner.cr), "cloneOnSpawnedThread should only be called on spawned threads");
243 /*! \brief The callback used for running on spawned threads.
245 * Obtains the pointer to the master mdrunner object from the one
246 * argument permitted to the thread-launch API call, copies it to make
247 * a new runner for this thread, reinitializes necessary data, and
248 * proceeds to the simulation. */
249 static void mdrunner_start_fn(const void *arg)
253 auto masterMdrunner = reinterpret_cast<const gmx::Mdrunner *>(arg);
254 /* copy the arg list to make sure that it's thread-local. This
255 doesn't copy pointed-to items, of course; fnm, cr and fplog
256 are reset in the call below, all others should be const. */
257 gmx::Mdrunner mdrunner = masterMdrunner->cloneOnSpawnedThread();
260 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
264 /*! \brief Start thread-MPI threads.
266 * Called by mdrunner() to start a specific number of threads
267 * (including the main thread) for thread-parallel runs. This in turn
268 * calls mdrunner() for each thread. All options are the same as for
270 t_commrec *Mdrunner::spawnThreads(int numThreadsToLaunch) const
273 /* first check whether we even need to start tMPI */
274 if (numThreadsToLaunch < 2)
280 /* now spawn new threads that start mdrunner_start_fn(), while
281 the main thread returns. Thread affinity is handled later. */
282 if (tMPI_Init_fn(TRUE, numThreadsToLaunch, TMPI_AFFINITY_NONE,
283 mdrunner_start_fn, static_cast<const void*>(this)) != TMPI_SUCCESS)
285 GMX_THROW(gmx::InternalError("Failed to spawn thread-MPI threads"));
288 threadMpiMdrunnerAccessBarrier();
290 GMX_UNUSED_VALUE(mdrunner_start_fn);
293 return reinitialize_commrec_for_this_thread(cr);
298 /*! \brief Initialize variables for Verlet scheme simulation */
299 static void prepare_verlet_scheme(FILE *fplog,
303 const gmx_mtop_t *mtop,
305 bool makeGpuPairList,
306 const gmx::CpuInfo &cpuinfo)
308 /* For NVE simulations, we will retain the initial list buffer */
309 if (EI_DYNAMICS(ir->eI) &&
310 ir->verletbuf_tol > 0 &&
311 !(EI_MD(ir->eI) && ir->etc == etcNO))
313 /* Update the Verlet buffer size for the current run setup */
315 /* Here we assume SIMD-enabled kernels are being used. But as currently
316 * calc_verlet_buffer_size gives the same results for 4x8 and 4x4
317 * and 4x2 gives a larger buffer than 4x4, this is ok.
319 ListSetupType listType = (makeGpuPairList ? ListSetupType::Gpu : ListSetupType::CpuSimdWhenSupported);
320 VerletbufListSetup listSetup = verletbufGetSafeListSetup(listType);
322 const real rlist_new =
323 calcVerletBufferSize(*mtop, det(box), *ir, ir->nstlist, ir->nstlist - 1, -1, listSetup);
325 if (rlist_new != ir->rlist)
327 if (fplog != nullptr)
329 fprintf(fplog, "\nChanging rlist from %g to %g for non-bonded %dx%d atom kernels\n\n",
330 ir->rlist, rlist_new,
331 listSetup.cluster_size_i, listSetup.cluster_size_j);
333 ir->rlist = rlist_new;
337 if (nstlist_cmdline > 0 && (!EI_DYNAMICS(ir->eI) || ir->verletbuf_tol <= 0))
339 gmx_fatal(FARGS, "Can not set nstlist without %s",
340 !EI_DYNAMICS(ir->eI) ? "dynamics" : "verlet-buffer-tolerance");
343 if (EI_DYNAMICS(ir->eI))
345 /* Set or try nstlist values */
346 increaseNstlist(fplog, cr, ir, nstlist_cmdline, mtop, box, makeGpuPairList, cpuinfo);
350 /*! \brief Override the nslist value in inputrec
352 * with value passed on the command line (if any)
354 static void override_nsteps_cmdline(const gmx::MDLogger &mdlog,
355 int64_t nsteps_cmdline,
360 /* override with anything else than the default -2 */
361 if (nsteps_cmdline > -2)
363 char sbuf_steps[STEPSTRSIZE];
364 char sbuf_msg[STRLEN];
366 ir->nsteps = nsteps_cmdline;
367 if (EI_DYNAMICS(ir->eI) && nsteps_cmdline != -1)
369 sprintf(sbuf_msg, "Overriding nsteps with value passed on the command line: %s steps, %.3g ps",
370 gmx_step_str(nsteps_cmdline, sbuf_steps),
371 fabs(nsteps_cmdline*ir->delta_t));
375 sprintf(sbuf_msg, "Overriding nsteps with value passed on the command line: %s steps",
376 gmx_step_str(nsteps_cmdline, sbuf_steps));
379 GMX_LOG(mdlog.warning).asParagraph().appendText(sbuf_msg);
381 else if (nsteps_cmdline < -2)
383 gmx_fatal(FARGS, "Invalid nsteps value passed on the command line: %" PRId64,
386 /* Do nothing if nsteps_cmdline == -2 */
392 /*! \brief Return whether GPU acceleration of nonbondeds is supported with the given settings.
394 * If not, and if a warning may be issued, logs a warning about
395 * falling back to CPU code. With thread-MPI, only the first
396 * call to this function should have \c issueWarning true. */
397 static bool gpuAccelerationOfNonbondedIsUseful(const MDLogger &mdlog,
398 const t_inputrec *ir,
401 if (ir->opts.ngener - ir->nwall > 1)
403 /* The GPU code does not support more than one energy group.
404 * If the user requested GPUs explicitly, a fatal error is given later.
408 GMX_LOG(mdlog.warning).asParagraph()
409 .appendText("Multiple energy groups is not implemented for GPUs, falling back to the CPU. "
410 "For better performance, run on the GPU without energy groups and then do "
411 "gmx mdrun -rerun option on the trajectory with an energy group .tpr file.");
418 //! Initializes the logger for mdrun.
419 static gmx::LoggerOwner buildLogger(FILE *fplog, const t_commrec *cr)
421 gmx::LoggerBuilder builder;
422 if (fplog != nullptr)
424 builder.addTargetFile(gmx::MDLogger::LogLevel::Info, fplog);
426 if (cr == nullptr || SIMMASTER(cr))
428 builder.addTargetStream(gmx::MDLogger::LogLevel::Warning,
429 &gmx::TextOutputFile::standardError());
431 return builder.build();
434 //! Make a TaskTarget from an mdrun argument string.
435 static TaskTarget findTaskTarget(const char *optionString)
437 TaskTarget returnValue = TaskTarget::Auto;
439 if (strncmp(optionString, "auto", 3) == 0)
441 returnValue = TaskTarget::Auto;
443 else if (strncmp(optionString, "cpu", 3) == 0)
445 returnValue = TaskTarget::Cpu;
447 else if (strncmp(optionString, "gpu", 3) == 0)
449 returnValue = TaskTarget::Gpu;
453 GMX_ASSERT(false, "Option string should have been checked for sanity already");
459 //! Finish run, aggregate data to print performance info.
460 static void finish_run(FILE *fplog,
461 const gmx::MDLogger &mdlog,
463 const t_inputrec *inputrec,
464 t_nrnb nrnb[], gmx_wallcycle_t wcycle,
465 gmx_walltime_accounting_t walltime_accounting,
466 nonbonded_verlet_t *nbv,
467 const gmx_pme_t *pme,
471 double nbfs = 0, mflop = 0;
473 elapsed_time_over_all_ranks,
474 elapsed_time_over_all_threads,
475 elapsed_time_over_all_threads_over_all_ranks;
476 /* Control whether it is valid to print a report. Only the
477 simulation master may print, but it should not do so if the run
478 terminated e.g. before a scheduled reset step. This is
479 complicated by the fact that PME ranks are unaware of the
480 reason why they were sent a pmerecvqxFINISH. To avoid
481 communication deadlocks, we always do the communication for the
482 report, even if we've decided not to write the report, because
483 how long it takes to finish the run is not important when we've
484 decided not to report on the simulation performance.
486 Further, we only report performance for dynamical integrators,
487 because those are the only ones for which we plan to
488 consider doing any optimizations. */
489 bool printReport = EI_DYNAMICS(inputrec->eI) && SIMMASTER(cr);
491 if (printReport && !walltime_accounting_get_valid_finish(walltime_accounting))
493 GMX_LOG(mdlog.warning).asParagraph().appendText("Simulation ended prematurely, no performance report will be written.");
498 std::unique_ptr<t_nrnb> nrnbTotalStorage;
501 nrnbTotalStorage = std::make_unique<t_nrnb>();
502 nrnb_tot = nrnbTotalStorage.get();
504 MPI_Allreduce(nrnb->n, nrnb_tot->n, eNRNB, MPI_DOUBLE, MPI_SUM,
513 elapsed_time = walltime_accounting_get_time_since_reset(walltime_accounting);
514 elapsed_time_over_all_threads = walltime_accounting_get_time_since_reset_over_all_threads(walltime_accounting);
518 /* reduce elapsed_time over all MPI ranks in the current simulation */
519 MPI_Allreduce(&elapsed_time,
520 &elapsed_time_over_all_ranks,
521 1, MPI_DOUBLE, MPI_SUM,
523 elapsed_time_over_all_ranks /= cr->nnodes;
524 /* Reduce elapsed_time_over_all_threads over all MPI ranks in the
525 * current simulation. */
526 MPI_Allreduce(&elapsed_time_over_all_threads,
527 &elapsed_time_over_all_threads_over_all_ranks,
528 1, MPI_DOUBLE, MPI_SUM,
534 elapsed_time_over_all_ranks = elapsed_time;
535 elapsed_time_over_all_threads_over_all_ranks = elapsed_time_over_all_threads;
540 print_flop(fplog, nrnb_tot, &nbfs, &mflop);
543 if (thisRankHasDuty(cr, DUTY_PP) && DOMAINDECOMP(cr))
545 print_dd_statistics(cr, inputrec, fplog);
548 /* TODO Move the responsibility for any scaling by thread counts
549 * to the code that handled the thread region, so that there's a
550 * mechanism to keep cycle counting working during the transition
551 * to task parallelism. */
552 int nthreads_pp = gmx_omp_nthreads_get(emntNonbonded);
553 int nthreads_pme = gmx_omp_nthreads_get(emntPME);
554 wallcycle_scale_by_num_threads(wcycle, thisRankHasDuty(cr, DUTY_PME) && !thisRankHasDuty(cr, DUTY_PP), nthreads_pp, nthreads_pme);
555 auto cycle_sum(wallcycle_sum(cr, wcycle));
559 auto nbnxn_gpu_timings = (nbv != nullptr && nbv->useGpu()) ? Nbnxm::gpu_get_timings(nbv->gpu_nbv) : nullptr;
560 gmx_wallclock_gpu_pme_t pme_gpu_timings = {};
562 if (pme_gpu_task_enabled(pme))
564 pme_gpu_get_timings(pme, &pme_gpu_timings);
566 wallcycle_print(fplog, mdlog, cr->nnodes, cr->npmenodes, nthreads_pp, nthreads_pme,
567 elapsed_time_over_all_ranks,
572 if (EI_DYNAMICS(inputrec->eI))
574 delta_t = inputrec->delta_t;
579 print_perf(fplog, elapsed_time_over_all_threads_over_all_ranks,
580 elapsed_time_over_all_ranks,
581 walltime_accounting_get_nsteps_done_since_reset(walltime_accounting),
582 delta_t, nbfs, mflop);
586 print_perf(stderr, elapsed_time_over_all_threads_over_all_ranks,
587 elapsed_time_over_all_ranks,
588 walltime_accounting_get_nsteps_done_since_reset(walltime_accounting),
589 delta_t, nbfs, mflop);
594 int Mdrunner::mdrunner()
597 t_forcerec *fr = nullptr;
598 t_fcdata *fcd = nullptr;
599 real ewaldcoeff_q = 0;
600 real ewaldcoeff_lj = 0;
601 int nChargePerturbed = -1, nTypePerturbed = 0;
602 gmx_wallcycle_t wcycle;
603 gmx_walltime_accounting_t walltime_accounting = nullptr;
604 gmx_membed_t * membed = nullptr;
605 gmx_hw_info_t *hwinfo = nullptr;
607 /* CAUTION: threads may be started later on in this function, so
608 cr doesn't reflect the final parallel state right now */
609 t_inputrec inputrecInstance;
610 t_inputrec *inputrec = &inputrecInstance;
613 bool doMembed = opt2bSet("-membed", filenames.size(), filenames.data());
614 bool doRerun = mdrunOptions.rerun;
616 // Handle task-assignment related user options.
617 EmulateGpuNonbonded emulateGpuNonbonded = (getenv("GMX_EMULATE_GPU") != nullptr ?
618 EmulateGpuNonbonded::Yes : EmulateGpuNonbonded::No);
620 std::vector<int> userGpuTaskAssignment;
623 userGpuTaskAssignment = parseUserTaskAssignmentString(hw_opt.userGpuTaskAssignment);
625 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
626 auto nonbondedTarget = findTaskTarget(nbpu_opt);
627 auto pmeTarget = findTaskTarget(pme_opt);
628 auto pmeFftTarget = findTaskTarget(pme_fft_opt);
629 auto bondedTarget = findTaskTarget(bonded_opt);
630 PmeRunMode pmeRunMode = PmeRunMode::None;
632 // Here we assume that SIMMASTER(cr) does not change even after the
633 // threads are started.
635 FILE *fplog = nullptr;
636 // If we are appending, we don't write log output because we need
637 // to check that the old log file matches what the checkpoint file
638 // expects. Otherwise, we should start to write log output now if
639 // there is a file ready for it.
640 if (logFileHandle != nullptr && startingBehavior != StartingBehavior::RestartWithAppending)
642 fplog = gmx_fio_getfp(logFileHandle);
644 gmx::LoggerOwner logOwner(buildLogger(fplog, cr));
645 gmx::MDLogger mdlog(logOwner.logger());
647 // report any development features that may be enabled by environment variables
648 reportDevelopmentFeatures(mdlog);
650 // With thread-MPI, the communicator changes after threads are
651 // launched, so this is rebuilt for the master rank at that
652 // time. The non-master ranks are fine to keep the one made here.
653 PhysicalNodeCommunicator physicalNodeComm(MPI_COMM_WORLD, gmx_physicalnode_id_hash());
654 hwinfo = gmx_detect_hardware(mdlog, physicalNodeComm);
656 gmx_print_detected_hardware(fplog, isMasterSimMasterRank(ms, MASTER(cr)), mdlog, hwinfo);
658 std::vector<int> gpuIdsToUse = makeGpuIdsToUse(hwinfo->gpu_info, hw_opt.gpuIdsAvailable);
660 // Print citation requests after all software/hardware printing
661 pleaseCiteGromacs(fplog);
663 std::unique_ptr<t_state> globalState;
667 /* Only the master rank has the global state */
668 globalState = std::make_unique<t_state>();
670 /* Read (nearly) all data required for the simulation */
671 read_tpx_state(ftp2fn(efTPR, filenames.size(), filenames.data()), inputrec, globalState.get(), &mtop);
674 /* Check and update the hardware options for internal consistency */
675 checkAndUpdateHardwareOptions(mdlog, &hw_opt, SIMMASTER(cr), domdecOptions.numPmeRanks);
677 if (GMX_THREAD_MPI && SIMMASTER(cr))
679 bool useGpuForNonbonded = false;
680 bool useGpuForPme = false;
683 // If the user specified the number of ranks, then we must
684 // respect that, but in default mode, we need to allow for
685 // the number of GPUs to choose the number of ranks.
686 auto canUseGpuForNonbonded = buildSupportsNonbondedOnGpu(nullptr);
687 useGpuForNonbonded = decideWhetherToUseGpusForNonbondedWithThreadMpi
688 (nonbondedTarget, gpuIdsToUse, userGpuTaskAssignment, emulateGpuNonbonded,
689 canUseGpuForNonbonded,
690 gpuAccelerationOfNonbondedIsUseful(mdlog, inputrec, GMX_THREAD_MPI),
691 hw_opt.nthreads_tmpi);
692 useGpuForPme = decideWhetherToUseGpusForPmeWithThreadMpi
693 (useGpuForNonbonded, pmeTarget, gpuIdsToUse, userGpuTaskAssignment,
694 *hwinfo, *inputrec, mtop, hw_opt.nthreads_tmpi, domdecOptions.numPmeRanks);
697 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
699 /* Determine how many thread-MPI ranks to start.
701 * TODO Over-writing the user-supplied value here does
702 * prevent any possible subsequent checks from working
704 hw_opt.nthreads_tmpi = get_nthreads_mpi(hwinfo,
713 // Now start the threads for thread MPI.
714 cr = spawnThreads(hw_opt.nthreads_tmpi);
715 /* The main thread continues here with a new cr. We don't deallocate
716 the old cr because other threads may still be reading it. */
717 // TODO Both master and spawned threads call dup_tfn and
718 // reinitialize_commrec_for_this_thread. Find a way to express
720 physicalNodeComm = PhysicalNodeCommunicator(MPI_COMM_WORLD, gmx_physicalnode_id_hash());
722 // END OF CAUTION: cr and physicalNodeComm are now reliable
726 /* now broadcast everything to the non-master nodes/threads: */
727 init_parallel(cr, inputrec, &mtop);
730 // Now each rank knows the inputrec that SIMMASTER read and used,
731 // and (if applicable) cr->nnodes has been assigned the number of
732 // thread-MPI ranks that have been chosen. The ranks can now all
733 // run the task-deciding functions and will agree on the result
734 // without needing to communicate.
736 // TODO Should we do the communication in debug mode to support
737 // having an assertion?
739 // Note that these variables describe only their own node.
741 // Note that when bonded interactions run on a GPU they always run
742 // alongside a nonbonded task, so do not influence task assignment
743 // even though they affect the force calculation workload.
744 bool useGpuForNonbonded = false;
745 bool useGpuForPme = false;
746 bool useGpuForBonded = false;
749 // It's possible that there are different numbers of GPUs on
750 // different nodes, which is the user's responsibilty to
751 // handle. If unsuitable, we will notice that during task
753 bool gpusWereDetected = hwinfo->ngpu_compatible_tot > 0;
754 auto canUseGpuForNonbonded = buildSupportsNonbondedOnGpu(nullptr);
755 useGpuForNonbonded = decideWhetherToUseGpusForNonbonded(nonbondedTarget, userGpuTaskAssignment,
757 canUseGpuForNonbonded,
758 gpuAccelerationOfNonbondedIsUseful(mdlog, inputrec, !GMX_THREAD_MPI),
760 useGpuForPme = decideWhetherToUseGpusForPme(useGpuForNonbonded, pmeTarget, userGpuTaskAssignment,
761 *hwinfo, *inputrec, mtop,
762 cr->nnodes, domdecOptions.numPmeRanks,
764 auto canUseGpuForBonded = buildSupportsGpuBondeds(nullptr) && inputSupportsGpuBondeds(*inputrec, mtop, nullptr);
766 decideWhetherToUseGpusForBonded(useGpuForNonbonded, useGpuForPme,
767 bondedTarget, canUseGpuForBonded,
768 EVDW_PME(inputrec->vdwtype),
769 EEL_PME_EWALD(inputrec->coulombtype),
770 domdecOptions.numPmeRanks, gpusWereDetected);
772 pmeRunMode = (useGpuForPme ? PmeRunMode::GPU : PmeRunMode::CPU);
773 if (pmeRunMode == PmeRunMode::GPU)
775 if (pmeFftTarget == TaskTarget::Cpu)
777 pmeRunMode = PmeRunMode::Mixed;
780 else if (pmeFftTarget == TaskTarget::Gpu)
782 gmx_fatal(FARGS, "Assigning FFTs to GPU requires PME to be assigned to GPU as well. With PME on CPU you should not be using -pmefft.");
785 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
788 // TODO: hide restraint implementation details from Mdrunner.
789 // There is nothing unique about restraints at this point as far as the
790 // Mdrunner is concerned. The Mdrunner should just be getting a sequence of
791 // factory functions from the SimulationContext on which to call mdModules_->add().
792 // TODO: capture all restraints into a single RestraintModule, passed to the runner builder.
793 for (auto && restraint : restraintManager_->getRestraints())
795 auto module = RestraintMDModule::create(restraint,
797 mdModules_->add(std::move(module));
800 // TODO: Error handling
801 mdModules_->assignOptionsToModules(*inputrec->params, nullptr);
802 if (inputrec->internalParameters != nullptr)
804 mdModules_->notifier().notify(*inputrec->internalParameters);
807 if (fplog != nullptr)
809 pr_inputrec(fplog, 0, "Input Parameters", inputrec, FALSE);
810 fprintf(fplog, "\n");
815 /* In rerun, set velocities to zero if present */
816 if (doRerun && ((globalState->flags & (1 << estV)) != 0))
818 // rerun does not use velocities
819 GMX_LOG(mdlog.info).asParagraph().appendText(
820 "Rerun trajectory contains velocities. Rerun does only evaluate "
821 "potential energy and forces. The velocities will be ignored.");
822 for (int i = 0; i < globalState->natoms; i++)
824 clear_rvec(globalState->v[i]);
826 globalState->flags &= ~(1 << estV);
829 /* now make sure the state is initialized and propagated */
830 set_state_entries(globalState.get(), inputrec);
833 /* NM and TPI parallelize over force/energy calculations, not atoms,
834 * so we need to initialize and broadcast the global state.
836 if (inputrec->eI == eiNM || inputrec->eI == eiTPI)
840 globalState = std::make_unique<t_state>();
842 broadcastStateWithoutDynamics(cr, globalState.get());
845 /* A parallel command line option consistency check that we can
846 only do after any threads have started. */
847 if (!PAR(cr) && (domdecOptions.numCells[XX] > 1 ||
848 domdecOptions.numCells[YY] > 1 ||
849 domdecOptions.numCells[ZZ] > 1 ||
850 domdecOptions.numPmeRanks > 0))
853 "The -dd or -npme option request a parallel simulation, "
855 "but %s was compiled without threads or MPI enabled", output_env_get_program_display_name(oenv));
858 "but the number of MPI-threads (option -ntmpi) is not set or is 1");
860 "but %s was not started through mpirun/mpiexec or only one rank was requested through mpirun/mpiexec", output_env_get_program_display_name(oenv));
866 (EI_ENERGY_MINIMIZATION(inputrec->eI) || eiNM == inputrec->eI))
868 gmx_fatal(FARGS, "The .mdp file specified an energy mininization or normal mode algorithm, and these are not compatible with mdrun -rerun");
871 if (!(EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype)))
873 if (domdecOptions.numPmeRanks > 0)
875 gmx_fatal_collective(FARGS, cr->mpi_comm_mysim, MASTER(cr),
876 "PME-only ranks are requested, but the system does not use PME for electrostatics or LJ");
879 domdecOptions.numPmeRanks = 0;
882 if (useGpuForNonbonded && domdecOptions.numPmeRanks < 0)
884 /* With NB GPUs we don't automatically use PME-only CPU ranks. PME ranks can
885 * improve performance with many threads per GPU, since our OpenMP
886 * scaling is bad, but it's difficult to automate the setup.
888 domdecOptions.numPmeRanks = 0;
892 if (domdecOptions.numPmeRanks < 0)
894 domdecOptions.numPmeRanks = 0;
895 // TODO possibly print a note that one can opt-in for a separate PME GPU rank?
899 GMX_RELEASE_ASSERT(domdecOptions.numPmeRanks <= 1, "PME GPU decomposition is not supported");
906 fcRegisterSteps(inputrec->nsteps, inputrec->init_step);
910 /* NMR restraints must be initialized before load_checkpoint,
911 * since with time averaging the history is added to t_state.
912 * For proper consistency check we therefore need to extend
914 * So the PME-only nodes (if present) will also initialize
915 * the distance restraints.
919 /* This needs to be called before read_checkpoint to extend the state */
920 init_disres(fplog, &mtop, inputrec, cr, ms, fcd, globalState.get(), replExParams.exchangeInterval > 0);
922 init_orires(fplog, &mtop, inputrec, cr, ms, globalState.get(), &(fcd->orires));
924 auto deform = prepareBoxDeformation(globalState->box, cr, *inputrec);
926 ObservablesHistory observablesHistory = {};
928 if (startingBehavior != StartingBehavior::NewSimulation)
930 /* Check if checkpoint file exists before doing continuation.
931 * This way we can use identical input options for the first and subsequent runs...
933 if (mdrunOptions.numStepsCommandline > -2)
935 /* Temporarily set the number of steps to unmlimited to avoid
936 * triggering the nsteps check in load_checkpoint().
937 * This hack will go away soon when the -nsteps option is removed.
939 inputrec->nsteps = -1;
942 load_checkpoint(opt2fn_master("-cpi", filenames.size(), filenames.data(), cr),
944 cr, domdecOptions.numCells,
945 inputrec, globalState.get(),
947 mdrunOptions.reproducible);
949 if (startingBehavior == StartingBehavior::RestartWithAppending && logFileHandle)
951 // Now we can start normal logging to the truncated log file.
952 fplog = gmx_fio_getfp(logFileHandle);
953 prepareLogAppending(fplog);
954 logOwner = buildLogger(fplog, cr);
955 mdlog = logOwner.logger();
959 if (mdrunOptions.numStepsCommandline > -2)
961 GMX_LOG(mdlog.info).asParagraph().
962 appendText("The -nsteps functionality is deprecated, and may be removed in a future version. "
963 "Consider using gmx convert-tpr -nsteps or changing the appropriate .mdp file field.");
965 /* override nsteps with value set on the commamdline */
966 override_nsteps_cmdline(mdlog, mdrunOptions.numStepsCommandline, inputrec);
970 copy_mat(globalState->box, box);
975 gmx_bcast(sizeof(box), box, cr);
978 if (inputrec->cutoff_scheme != ecutsVERLET)
980 gmx_fatal(FARGS, "This group-scheme .tpr file can no longer be run by mdrun. Please update to the Verlet scheme, or use an earlier version of GROMACS if necessary.");
982 /* Update rlist and nstlist. */
983 prepare_verlet_scheme(fplog, cr, inputrec, nstlist_cmdline, &mtop, box,
984 useGpuForNonbonded || (emulateGpuNonbonded == EmulateGpuNonbonded::Yes), *hwinfo->cpuInfo);
986 LocalAtomSetManager atomSets;
988 if (PAR(cr) && !(EI_TPI(inputrec->eI) ||
989 inputrec->eI == eiNM))
991 cr->dd = init_domain_decomposition(mdlog, cr, domdecOptions, mdrunOptions,
993 box, positionsFromStatePointer(globalState.get()),
995 mdModules_->notifier().notify(&atomSets);
996 // Note that local state still does not exist yet.
1000 /* PME, if used, is done on all nodes with 1D decomposition */
1002 cr->duty = (DUTY_PP | DUTY_PME);
1004 if (inputrec->ePBC == epbcSCREW)
1007 "pbc=screw is only implemented with domain decomposition");
1013 /* After possible communicator splitting in make_dd_communicators.
1014 * we can set up the intra/inter node communication.
1016 gmx_setup_nodecomm(fplog, cr);
1022 GMX_LOG(mdlog.warning).asParagraph().appendTextFormatted(
1023 "This is simulation %d out of %d running as a composite GROMACS\n"
1024 "multi-simulation job. Setup for this simulation:\n",
1027 GMX_LOG(mdlog.warning).appendTextFormatted(
1028 "Using %d MPI %s\n",
1031 cr->nnodes == 1 ? "thread" : "threads"
1033 cr->nnodes == 1 ? "process" : "processes"
1039 // If mdrun -pin auto honors any affinity setting that already
1040 // exists. If so, it is nice to provide feedback about whether
1041 // that existing affinity setting was from OpenMP or something
1042 // else, so we run this code both before and after we initialize
1043 // the OpenMP support.
1044 gmx_check_thread_affinity_set(mdlog,
1045 &hw_opt, hwinfo->nthreads_hw_avail, FALSE);
1046 /* Check and update the number of OpenMP threads requested */
1047 checkAndUpdateRequestedNumOpenmpThreads(&hw_opt, *hwinfo, cr, ms, physicalNodeComm.size_,
1050 gmx_omp_nthreads_init(mdlog, cr,
1051 hwinfo->nthreads_hw_avail,
1052 physicalNodeComm.size_,
1053 hw_opt.nthreads_omp,
1054 hw_opt.nthreads_omp_pme,
1055 !thisRankHasDuty(cr, DUTY_PP));
1057 // Enable FP exception detection, but not in
1058 // Release mode and not for compilers with known buggy FP
1059 // exception support (clang with any optimization) or suspected
1060 // buggy FP exception support (gcc 7.* with optimization).
1061 #if !defined NDEBUG && \
1062 !((defined __clang__ || (defined(__GNUC__) && !defined(__ICC) && __GNUC__ == 7)) \
1063 && defined __OPTIMIZE__)
1064 const bool bEnableFPE = true;
1066 const bool bEnableFPE = false;
1068 //FIXME - reconcile with gmx_feenableexcept() call from CommandLineModuleManager::run()
1071 gmx_feenableexcept();
1074 // Build a data structure that expresses which kinds of non-bonded
1075 // task are handled by this rank.
1077 // TODO Later, this might become a loop over all registered modules
1078 // relevant to the mdp inputs, to find those that have such tasks.
1080 // TODO This could move before init_domain_decomposition() as part
1081 // of refactoring that separates the responsibility for duty
1082 // assignment from setup for communication between tasks, and
1083 // setup for tasks handled with a domain (ie including short-ranged
1084 // tasks, bonded tasks, etc.).
1086 // Note that in general useGpuForNonbonded, etc. can have a value
1087 // that is inconsistent with the presence of actual GPUs on any
1088 // rank, and that is not known to be a problem until the
1089 // duty of the ranks on a node become known.
1091 // TODO Later we might need the concept of computeTasksOnThisRank,
1092 // from which we construct gpuTasksOnThisRank.
1094 // Currently the DD code assigns duty to ranks that can
1095 // include PP work that currently can be executed on a single
1096 // GPU, if present and compatible. This has to be coordinated
1097 // across PP ranks on a node, with possible multiple devices
1098 // or sharing devices on a node, either from the user
1099 // selection, or automatically.
1100 auto haveGpus = !gpuIdsToUse.empty();
1101 std::vector<GpuTask> gpuTasksOnThisRank;
1102 if (thisRankHasDuty(cr, DUTY_PP))
1104 if (useGpuForNonbonded)
1106 // Note that any bonded tasks on a GPU always accompany a
1110 gpuTasksOnThisRank.push_back(GpuTask::Nonbonded);
1112 else if (nonbondedTarget == TaskTarget::Gpu)
1114 gmx_fatal(FARGS, "Cannot run short-ranged nonbonded interactions on a GPU because no GPU is detected.");
1116 else if (bondedTarget == TaskTarget::Gpu)
1118 gmx_fatal(FARGS, "Cannot run bonded interactions on a GPU because no GPU is detected.");
1122 // TODO cr->duty & DUTY_PME should imply that a PME algorithm is active, but currently does not.
1123 if (EEL_PME(inputrec->coulombtype) && (thisRankHasDuty(cr, DUTY_PME)))
1129 gpuTasksOnThisRank.push_back(GpuTask::Pme);
1131 else if (pmeTarget == TaskTarget::Gpu)
1133 gmx_fatal(FARGS, "Cannot run PME on a GPU because no GPU is detected.");
1138 GpuTaskAssignment gpuTaskAssignment;
1141 // Produce the task assignment for this rank.
1142 gpuTaskAssignment = runTaskAssignment(gpuIdsToUse, userGpuTaskAssignment, *hwinfo,
1143 mdlog, cr, ms, physicalNodeComm, gpuTasksOnThisRank,
1144 useGpuForBonded, pmeRunMode);
1146 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1148 /* Prevent other ranks from continuing after an issue was found
1149 * and reported as a fatal error.
1151 * TODO This function implements a barrier so that MPI runtimes
1152 * can organize an orderly shutdown if one of the ranks has had to
1153 * issue a fatal error in various code already run. When we have
1154 * MPI-aware error handling and reporting, this should be
1159 MPI_Barrier(cr->mpi_comm_mysim);
1165 MPI_Barrier(ms->mpi_comm_masters);
1167 /* We need another barrier to prevent non-master ranks from contiuing
1168 * when an error occured in a different simulation.
1170 MPI_Barrier(cr->mpi_comm_mysim);
1174 /* Now that we know the setup is consistent, check for efficiency */
1175 check_resource_division_efficiency(hwinfo, !gpuTaskAssignment.empty(), mdrunOptions.ntompOptionIsSet,
1178 gmx_device_info_t *nonbondedDeviceInfo = nullptr;
1180 if (thisRankHasDuty(cr, DUTY_PP))
1182 // This works because only one task of each type is currently permitted.
1183 auto nbGpuTaskMapping = std::find_if(gpuTaskAssignment.begin(), gpuTaskAssignment.end(),
1184 hasTaskType<GpuTask::Nonbonded>);
1185 if (nbGpuTaskMapping != gpuTaskAssignment.end())
1187 int nonbondedDeviceId = nbGpuTaskMapping->deviceId_;
1188 nonbondedDeviceInfo = getDeviceInfo(hwinfo->gpu_info, nonbondedDeviceId);
1189 init_gpu(nonbondedDeviceInfo);
1191 if (DOMAINDECOMP(cr))
1193 /* When we share GPUs over ranks, we need to know this for the DLB */
1194 dd_setup_dlb_resource_sharing(cr, nonbondedDeviceId);
1200 std::unique_ptr<ClfftInitializer> initializedClfftLibrary;
1202 gmx_device_info_t *pmeDeviceInfo = nullptr;
1203 // Later, this program could contain kernels that might be later
1204 // re-used as auto-tuning progresses, or subsequent simulations
1206 PmeGpuProgramStorage pmeGpuProgram;
1207 // This works because only one task of each type is currently permitted.
1208 auto pmeGpuTaskMapping = std::find_if(gpuTaskAssignment.begin(), gpuTaskAssignment.end(), hasTaskType<GpuTask::Pme>);
1209 const bool thisRankHasPmeGpuTask = (pmeGpuTaskMapping != gpuTaskAssignment.end());
1210 if (thisRankHasPmeGpuTask)
1212 pmeDeviceInfo = getDeviceInfo(hwinfo->gpu_info, pmeGpuTaskMapping->deviceId_);
1213 init_gpu(pmeDeviceInfo);
1214 pmeGpuProgram = buildPmeGpuProgram(pmeDeviceInfo);
1215 // TODO It would be nice to move this logic into the factory
1216 // function. See Redmine #2535.
1217 bool isMasterThread = !GMX_THREAD_MPI || MASTER(cr);
1218 if (pmeRunMode == PmeRunMode::GPU && !initializedClfftLibrary && isMasterThread)
1220 initializedClfftLibrary = initializeClfftLibrary();
1224 /* getting number of PP/PME threads on this MPI / tMPI rank.
1225 PME: env variable should be read only on one node to make sure it is
1226 identical everywhere;
1228 const int numThreadsOnThisRank =
1229 thisRankHasDuty(cr, DUTY_PP) ? gmx_omp_nthreads_get(emntNonbonded) : gmx_omp_nthreads_get(emntPME);
1230 checkHardwareOversubscription(numThreadsOnThisRank, cr->nodeid,
1231 *hwinfo->hardwareTopology,
1232 physicalNodeComm, mdlog);
1234 if (hw_opt.threadAffinity != ThreadAffinity::Off)
1236 /* Before setting affinity, check whether the affinity has changed
1237 * - which indicates that probably the OpenMP library has changed it
1238 * since we first checked).
1240 gmx_check_thread_affinity_set(mdlog,
1241 &hw_opt, hwinfo->nthreads_hw_avail, TRUE);
1243 int numThreadsOnThisNode, intraNodeThreadOffset;
1244 analyzeThreadsOnThisNode(physicalNodeComm, numThreadsOnThisRank, &numThreadsOnThisNode,
1245 &intraNodeThreadOffset);
1247 /* Set the CPU affinity */
1248 gmx_set_thread_affinity(mdlog, cr, &hw_opt, *hwinfo->hardwareTopology,
1249 numThreadsOnThisRank, numThreadsOnThisNode,
1250 intraNodeThreadOffset, nullptr);
1253 if (mdrunOptions.timingOptions.resetStep > -1)
1255 GMX_LOG(mdlog.info).asParagraph().
1256 appendText("The -resetstep functionality is deprecated, and may be removed in a future version.");
1258 wcycle = wallcycle_init(fplog, mdrunOptions.timingOptions.resetStep, cr);
1262 /* Master synchronizes its value of reset_counters with all nodes
1263 * including PME only nodes */
1264 int64_t reset_counters = wcycle_get_reset_counters(wcycle);
1265 gmx_bcast_sim(sizeof(reset_counters), &reset_counters, cr);
1266 wcycle_set_reset_counters(wcycle, reset_counters);
1269 // Membrane embedding must be initialized before we call init_forcerec()
1274 fprintf(stderr, "Initializing membed");
1276 /* Note that membed cannot work in parallel because mtop is
1277 * changed here. Fix this if we ever want to make it run with
1278 * multiple ranks. */
1279 membed = init_membed(fplog, filenames.size(), filenames.data(), &mtop, inputrec, globalState.get(), cr,
1281 .checkpointOptions.period);
1284 std::unique_ptr<MDAtoms> mdAtoms;
1285 std::unique_ptr<gmx_vsite_t> vsite;
1288 if (thisRankHasDuty(cr, DUTY_PP))
1290 mdModules_->notifier().notify(CommunicationIsSetup {*cr});
1291 /* Initiate forcerecord */
1292 fr = new t_forcerec;
1293 fr->forceProviders = mdModules_->initForceProviders();
1294 init_forcerec(fplog, mdlog, fr, fcd,
1295 inputrec, &mtop, cr, box,
1296 opt2fn("-table", filenames.size(), filenames.data()),
1297 opt2fn("-tablep", filenames.size(), filenames.data()),
1298 opt2fns("-tableb", filenames.size(), filenames.data()),
1299 *hwinfo, nonbondedDeviceInfo,
1305 /* Initialize the mdAtoms structure.
1306 * mdAtoms is not filled with atom data,
1307 * as this can not be done now with domain decomposition.
1309 mdAtoms = makeMDAtoms(fplog, mtop, *inputrec, thisRankHasPmeGpuTask);
1310 if (globalState && thisRankHasPmeGpuTask)
1312 // The pinning of coordinates in the global state object works, because we only use
1313 // PME on GPU without DD or on a separate PME rank, and because the local state pointer
1314 // points to the global state object without DD.
1315 // FIXME: MD and EM separately set up the local state - this should happen in the same function,
1316 // which should also perform the pinning.
1317 changePinningPolicy(&globalState->x, pme_get_pinning_policy());
1320 /* Initialize the virtual site communication */
1321 vsite = initVsite(mtop, cr);
1323 calc_shifts(box, fr->shift_vec);
1325 /* With periodic molecules the charge groups should be whole at start up
1326 * and the virtual sites should not be far from their proper positions.
1328 if (!inputrec->bContinuation && MASTER(cr) &&
1329 !(inputrec->ePBC != epbcNONE && inputrec->bPeriodicMols))
1331 /* Make molecules whole at start of run */
1332 if (fr->ePBC != epbcNONE)
1334 do_pbc_first_mtop(fplog, inputrec->ePBC, box, &mtop, globalState->x.rvec_array());
1338 /* Correct initial vsite positions are required
1339 * for the initial distribution in the domain decomposition
1340 * and for the initial shell prediction.
1342 constructVsitesGlobal(mtop, globalState->x);
1346 if (EEL_PME(fr->ic->eeltype) || EVDW_PME(fr->ic->vdwtype))
1348 ewaldcoeff_q = fr->ic->ewaldcoeff_q;
1349 ewaldcoeff_lj = fr->ic->ewaldcoeff_lj;
1354 /* This is a PME only node */
1356 GMX_ASSERT(globalState == nullptr, "We don't need the state on a PME only rank and expect it to be unitialized");
1358 ewaldcoeff_q = calc_ewaldcoeff_q(inputrec->rcoulomb, inputrec->ewald_rtol);
1359 ewaldcoeff_lj = calc_ewaldcoeff_lj(inputrec->rvdw, inputrec->ewald_rtol_lj);
1362 gmx_pme_t *sepPmeData = nullptr;
1363 // This reference hides the fact that PME data is owned by runner on PME-only ranks and by forcerec on other ranks
1364 GMX_ASSERT(thisRankHasDuty(cr, DUTY_PP) == (fr != nullptr), "Double-checking that only PME-only ranks have no forcerec");
1365 gmx_pme_t * &pmedata = fr ? fr->pmedata : sepPmeData;
1367 /* Initiate PME if necessary,
1368 * either on all nodes or on dedicated PME nodes only. */
1369 if (EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype))
1371 if (mdAtoms && mdAtoms->mdatoms())
1373 nChargePerturbed = mdAtoms->mdatoms()->nChargePerturbed;
1374 if (EVDW_PME(inputrec->vdwtype))
1376 nTypePerturbed = mdAtoms->mdatoms()->nTypePerturbed;
1379 if (cr->npmenodes > 0)
1381 /* The PME only nodes need to know nChargePerturbed(FEP on Q) and nTypePerturbed(FEP on LJ)*/
1382 gmx_bcast_sim(sizeof(nChargePerturbed), &nChargePerturbed, cr);
1383 gmx_bcast_sim(sizeof(nTypePerturbed), &nTypePerturbed, cr);
1386 if (thisRankHasDuty(cr, DUTY_PME))
1390 pmedata = gmx_pme_init(cr,
1391 getNumPmeDomains(cr->dd),
1393 nChargePerturbed != 0, nTypePerturbed != 0,
1394 mdrunOptions.reproducible,
1395 ewaldcoeff_q, ewaldcoeff_lj,
1396 gmx_omp_nthreads_get(emntPME),
1397 pmeRunMode, nullptr,
1398 pmeDeviceInfo, pmeGpuProgram.get(), mdlog);
1400 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1405 if (EI_DYNAMICS(inputrec->eI))
1407 /* Turn on signal handling on all nodes */
1409 * (A user signal from the PME nodes (if any)
1410 * is communicated to the PP nodes.
1412 signal_handler_install();
1415 pull_t *pull_work = nullptr;
1416 if (thisRankHasDuty(cr, DUTY_PP))
1418 /* Assumes uniform use of the number of OpenMP threads */
1419 walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(emntDefault));
1421 if (inputrec->bPull)
1423 /* Initialize pull code */
1425 init_pull(fplog, inputrec->pull, inputrec,
1426 &mtop, cr, &atomSets, inputrec->fepvals->init_lambda);
1427 if (inputrec->pull->bXOutAverage || inputrec->pull->bFOutAverage)
1429 initPullHistory(pull_work, &observablesHistory);
1431 if (EI_DYNAMICS(inputrec->eI) && MASTER(cr))
1433 init_pull_output_files(pull_work,
1434 filenames.size(), filenames.data(), oenv,
1439 std::unique_ptr<EnforcedRotation> enforcedRotation;
1442 /* Initialize enforced rotation code */
1443 enforcedRotation = init_rot(fplog,
1456 t_swap *swap = nullptr;
1457 if (inputrec->eSwapCoords != eswapNO)
1459 /* Initialize ion swapping code */
1460 swap = init_swapcoords(fplog, inputrec,
1461 opt2fn_master("-swap", filenames.size(), filenames.data(), cr),
1462 &mtop, globalState.get(), &observablesHistory,
1463 cr, &atomSets, oenv, mdrunOptions,
1467 /* Let makeConstraints know whether we have essential dynamics constraints.
1468 * TODO: inputrec should tell us whether we use an algorithm, not a file option or the checkpoint
1470 bool doEssentialDynamics = (opt2fn_null("-ei", filenames.size(), filenames.data()) != nullptr
1471 || observablesHistory.edsamHistory);
1472 auto constr = makeConstraints(mtop, *inputrec, pull_work, doEssentialDynamics,
1473 fplog, *mdAtoms->mdatoms(),
1474 cr, ms, &nrnb, wcycle, fr->bMolPBC);
1476 /* Energy terms and groups */
1477 gmx_enerdata_t enerd(mtop.groups.groups[SimulationAtomGroupType::EnergyOutput].size(), inputrec->fepvals->n_lambda);
1479 /* Kinetic energy data */
1480 gmx_ekindata_t ekind;
1481 init_ekindata(fplog, &mtop, &(inputrec->opts), &ekind, inputrec->cos_accel);
1483 /* Set up interactive MD (IMD) */
1484 auto imdSession = makeImdSession(inputrec, cr, wcycle, &enerd, ms, &mtop, mdlog,
1485 MASTER(cr) ? globalState->x.rvec_array() : nullptr,
1486 filenames.size(), filenames.data(), oenv, mdrunOptions.imdOptions,
1489 if (DOMAINDECOMP(cr))
1491 GMX_RELEASE_ASSERT(fr, "fr was NULL while cr->duty was DUTY_PP");
1492 /* This call is not included in init_domain_decomposition mainly
1493 * because fr->cginfo_mb is set later.
1495 dd_init_bondeds(fplog, cr->dd, &mtop, vsite.get(), inputrec,
1496 domdecOptions.checkBondedInteractions,
1500 // TODO This is not the right place to manage the lifetime of
1501 // this data structure, but currently it's the easiest way to
1502 // make it work. Later, it should probably be made/updated
1503 // after the workload for the lifetime of a PP domain is
1505 PpForceWorkload ppForceWorkload;
1507 GMX_ASSERT(stopHandlerBuilder_, "Runner must provide StopHandlerBuilder to simulator.");
1508 SimulatorBuilder simulatorBuilder;
1510 // build and run simulator object based on user-input
1511 auto simulator = simulatorBuilder.build(
1512 fplog, cr, ms, mdlog, static_cast<int>(filenames.size()), filenames.data(),
1516 vsite.get(), constr.get(),
1517 enforcedRotation ? enforcedRotation->getLegacyEnfrot() : nullptr,
1519 mdModules_->outputProvider(),
1520 inputrec, imdSession.get(), pull_work, swap, &mtop,
1523 &observablesHistory,
1524 mdAtoms.get(), &nrnb, wcycle, fr,
1530 walltime_accounting,
1531 std::move(stopHandlerBuilder_),
1535 if (inputrec->bPull)
1537 finish_pull(pull_work);
1539 finish_swapcoords(swap);
1543 GMX_RELEASE_ASSERT(pmedata, "pmedata was NULL while cr->duty was not DUTY_PP");
1545 walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(emntPME));
1546 gmx_pmeonly(pmedata, cr, &nrnb, wcycle, walltime_accounting, inputrec, pmeRunMode);
1549 wallcycle_stop(wcycle, ewcRUN);
1551 /* Finish up, write some stuff
1552 * if rerunMD, don't write last frame again
1554 finish_run(fplog, mdlog, cr,
1555 inputrec, &nrnb, wcycle, walltime_accounting,
1556 fr ? fr->nbv.get() : nullptr,
1558 EI_DYNAMICS(inputrec->eI) && !isMultiSim(ms));
1560 // clean up cycle counter
1561 wallcycle_destroy(wcycle);
1566 gmx_pme_destroy(pmedata);
1570 // FIXME: this is only here to manually unpin mdAtoms->chargeA_ and state->x,
1571 // before we destroy the GPU context(s) in free_gpu_resources().
1572 // Pinned buffers are associated with contexts in CUDA.
1573 // As soon as we destroy GPU contexts after mdrunner() exits, these lines should go.
1574 mdAtoms.reset(nullptr);
1575 globalState.reset(nullptr);
1576 mdModules_.reset(nullptr); // destruct force providers here as they might also use the GPU
1578 /* Free GPU memory and set a physical node tMPI barrier (which should eventually go away) */
1579 free_gpu_resources(fr, physicalNodeComm, hwinfo->gpu_info);
1580 free_gpu(nonbondedDeviceInfo);
1581 free_gpu(pmeDeviceInfo);
1582 done_forcerec(fr, mtop.molblock.size());
1587 free_membed(membed);
1590 /* Does what it says */
1591 print_date_and_time(fplog, cr->nodeid, "Finished mdrun", gmx_gettime());
1592 walltime_accounting_destroy(walltime_accounting);
1594 // Ensure log file content is written
1597 gmx_fio_flush(logFileHandle);
1600 /* Reset FPEs (important for unit tests) by disabling them. Assumes no
1601 * exceptions were enabled before function was called. */
1604 gmx_fedisableexcept();
1607 auto rc = static_cast<int>(gmx_get_stop_condition());
1610 /* we need to join all threads. The sub-threads join when they
1611 exit this function, but the master thread needs to be told to
1613 if (PAR(cr) && MASTER(cr))
1617 //TODO free commrec in MPI simulations
1623 Mdrunner::~Mdrunner()
1625 // Clean up of the Manager.
1626 // This will end up getting called on every thread-MPI rank, which is unnecessary,
1627 // but okay as long as threads synchronize some time before adding or accessing
1628 // a new set of restraints.
1629 if (restraintManager_)
1631 restraintManager_->clear();
1632 GMX_ASSERT(restraintManager_->countRestraints() == 0,
1633 "restraints added during runner life time should be cleared at runner destruction.");
1637 void Mdrunner::addPotential(std::shared_ptr<gmx::IRestraintPotential> puller,
1640 GMX_ASSERT(restraintManager_, "Mdrunner must have a restraint manager.");
1641 // Not sure if this should be logged through the md logger or something else,
1642 // but it is helpful to have some sort of INFO level message sent somewhere.
1643 // std::cout << "Registering restraint named " << name << std::endl;
1645 // When multiple restraints are used, it may be wasteful to register them separately.
1646 // Maybe instead register an entire Restraint Manager as a force provider.
1647 restraintManager_->addToSpec(std::move(puller),
1651 Mdrunner::Mdrunner(std::unique_ptr<MDModules> mdModules)
1652 : mdModules_(std::move(mdModules))
1656 Mdrunner::Mdrunner(Mdrunner &&) noexcept = default;
1658 //NOLINTNEXTLINE(performance-noexcept-move-constructor) working around GCC bug 58265
1659 Mdrunner &Mdrunner::operator=(Mdrunner && /*handle*/) noexcept(BUGFREE_NOEXCEPT_STRING) = default;
1661 class Mdrunner::BuilderImplementation
1664 BuilderImplementation() = delete;
1665 BuilderImplementation(std::unique_ptr<MDModules> mdModules,
1666 SimulationContext * context);
1667 ~BuilderImplementation();
1669 BuilderImplementation &setExtraMdrunOptions(const MdrunOptions &options,
1670 real forceWarningThreshold,
1671 StartingBehavior startingBehavior);
1673 void addDomdec(const DomdecOptions &options);
1675 void addVerletList(int nstlist);
1677 void addReplicaExchange(const ReplicaExchangeParameters ¶ms);
1679 void addMultiSim(gmx_multisim_t* multisim);
1681 void addNonBonded(const char* nbpu_opt);
1683 void addPME(const char* pme_opt_, const char* pme_fft_opt_);
1685 void addBondedTaskAssignment(const char* bonded_opt);
1687 void addHardwareOptions(const gmx_hw_opt_t &hardwareOptions);
1689 void addFilenames(ArrayRef <const t_filenm> filenames);
1691 void addOutputEnvironment(gmx_output_env_t* outputEnvironment);
1693 void addLogFile(t_fileio *logFileHandle);
1695 void addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder);
1701 // Default parameters copied from runner.h
1702 // \todo Clarify source(s) of default parameters.
1704 const char* nbpu_opt_ = nullptr;
1705 const char* pme_opt_ = nullptr;
1706 const char* pme_fft_opt_ = nullptr;
1707 const char *bonded_opt_ = nullptr;
1709 MdrunOptions mdrunOptions_;
1711 DomdecOptions domdecOptions_;
1713 ReplicaExchangeParameters replicaExchangeParameters_;
1715 //! Command-line override for the duration of a neighbor list with the Verlet scheme.
1718 //! Non-owning multisim communicator handle.
1719 std::unique_ptr<gmx_multisim_t*> multisim_ = nullptr;
1721 //! Print a warning if any force is larger than this (in kJ/mol nm).
1722 real forceWarningThreshold_ = -1;
1724 //! Whether the simulation will start afresh, or restart with/without appending.
1725 StartingBehavior startingBehavior_ = StartingBehavior::NewSimulation;
1727 //! The modules that comprise the functionality of mdrun.
1728 std::unique_ptr<MDModules> mdModules_;
1730 /*! \brief Non-owning pointer to SimulationContext (owned and managed by client)
1733 * \todo Establish robust protocol to make sure resources remain valid.
1734 * SimulationContext will likely be separated into multiple layers for
1735 * different levels of access and different phases of execution. Ref
1736 * https://redmine.gromacs.org/issues/2375
1737 * https://redmine.gromacs.org/issues/2587
1739 SimulationContext* context_ = nullptr;
1741 //! \brief Parallelism information.
1742 gmx_hw_opt_t hardwareOptions_;
1744 //! filename options for simulation.
1745 ArrayRef<const t_filenm> filenames_;
1747 /*! \brief Handle to output environment.
1749 * \todo gmx_output_env_t needs lifetime management.
1751 gmx_output_env_t* outputEnvironment_ = nullptr;
1753 /*! \brief Non-owning handle to MD log file.
1755 * \todo Context should own output facilities for client.
1756 * \todo Improve log file handle management.
1758 * Code managing the FILE* relies on the ability to set it to
1759 * nullptr to check whether the filehandle is valid.
1761 t_fileio* logFileHandle_ = nullptr;
1764 * \brief Builder for simulation stop signal handler.
1766 std::unique_ptr<StopHandlerBuilder> stopHandlerBuilder_ = nullptr;
1769 Mdrunner::BuilderImplementation::BuilderImplementation(std::unique_ptr<MDModules> mdModules,
1770 SimulationContext * context) :
1771 mdModules_(std::move(mdModules)),
1774 GMX_ASSERT(context_, "Bug found. It should not be possible to construct builder without a valid context.");
1777 Mdrunner::BuilderImplementation::~BuilderImplementation() = default;
1779 Mdrunner::BuilderImplementation &
1780 Mdrunner::BuilderImplementation::setExtraMdrunOptions(const MdrunOptions &options,
1781 const real forceWarningThreshold,
1782 const StartingBehavior startingBehavior)
1784 mdrunOptions_ = options;
1785 forceWarningThreshold_ = forceWarningThreshold;
1786 startingBehavior_ = startingBehavior;
1790 void Mdrunner::BuilderImplementation::addDomdec(const DomdecOptions &options)
1792 domdecOptions_ = options;
1795 void Mdrunner::BuilderImplementation::addVerletList(int nstlist)
1800 void Mdrunner::BuilderImplementation::addReplicaExchange(const ReplicaExchangeParameters ¶ms)
1802 replicaExchangeParameters_ = params;
1805 void Mdrunner::BuilderImplementation::addMultiSim(gmx_multisim_t* multisim)
1807 multisim_ = std::make_unique<gmx_multisim_t*>(multisim);
1810 Mdrunner Mdrunner::BuilderImplementation::build()
1812 auto newRunner = Mdrunner(std::move(mdModules_));
1814 GMX_ASSERT(context_, "Bug found. It should not be possible to call build() without a valid context.");
1816 newRunner.mdrunOptions = mdrunOptions_;
1817 newRunner.pforce = forceWarningThreshold_;
1818 newRunner.startingBehavior = startingBehavior_;
1819 newRunner.domdecOptions = domdecOptions_;
1821 // \todo determine an invariant to check or confirm that all gmx_hw_opt_t objects are valid
1822 newRunner.hw_opt = hardwareOptions_;
1824 // No invariant to check. This parameter exists to optionally override other behavior.
1825 newRunner.nstlist_cmdline = nstlist_;
1827 newRunner.replExParams = replicaExchangeParameters_;
1829 newRunner.filenames = filenames_;
1831 GMX_ASSERT(context_->communicationRecord_, "SimulationContext communications not initialized.");
1832 newRunner.cr = context_->communicationRecord_;
1836 // nullptr is a valid value for the multisim handle, so we don't check the pointed-to pointer.
1837 newRunner.ms = *multisim_;
1841 GMX_THROW(gmx::APIError("MdrunnerBuilder::addMultiSim() is required before build()"));
1844 // \todo Clarify ownership and lifetime management for gmx_output_env_t
1845 // \todo Update sanity checking when output environment has clearly specified invariants.
1846 // Initialization and default values for oenv are not well specified in the current version.
1847 if (outputEnvironment_)
1849 newRunner.oenv = outputEnvironment_;
1853 GMX_THROW(gmx::APIError("MdrunnerBuilder::addOutputEnvironment() is required before build()"));
1856 newRunner.logFileHandle = logFileHandle_;
1860 newRunner.nbpu_opt = nbpu_opt_;
1864 GMX_THROW(gmx::APIError("MdrunnerBuilder::addNonBonded() is required before build()"));
1867 if (pme_opt_ && pme_fft_opt_)
1869 newRunner.pme_opt = pme_opt_;
1870 newRunner.pme_fft_opt = pme_fft_opt_;
1874 GMX_THROW(gmx::APIError("MdrunnerBuilder::addElectrostatics() is required before build()"));
1879 newRunner.bonded_opt = bonded_opt_;
1883 GMX_THROW(gmx::APIError("MdrunnerBuilder::addBondedTaskAssignment() is required before build()"));
1886 newRunner.restraintManager_ = std::make_unique<gmx::RestraintManager>();
1888 if (stopHandlerBuilder_)
1890 newRunner.stopHandlerBuilder_ = std::move(stopHandlerBuilder_);
1894 newRunner.stopHandlerBuilder_ = std::make_unique<StopHandlerBuilder>();
1900 void Mdrunner::BuilderImplementation::addNonBonded(const char* nbpu_opt)
1902 nbpu_opt_ = nbpu_opt;
1905 void Mdrunner::BuilderImplementation::addPME(const char* pme_opt,
1906 const char* pme_fft_opt)
1909 pme_fft_opt_ = pme_fft_opt;
1912 void Mdrunner::BuilderImplementation::addBondedTaskAssignment(const char* bonded_opt)
1914 bonded_opt_ = bonded_opt;
1917 void Mdrunner::BuilderImplementation::addHardwareOptions(const gmx_hw_opt_t &hardwareOptions)
1919 hardwareOptions_ = hardwareOptions;
1922 void Mdrunner::BuilderImplementation::addFilenames(ArrayRef<const t_filenm> filenames)
1924 filenames_ = filenames;
1927 void Mdrunner::BuilderImplementation::addOutputEnvironment(gmx_output_env_t* outputEnvironment)
1929 outputEnvironment_ = outputEnvironment;
1932 void Mdrunner::BuilderImplementation::addLogFile(t_fileio *logFileHandle)
1934 logFileHandle_ = logFileHandle;
1937 void Mdrunner::BuilderImplementation::addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder)
1939 stopHandlerBuilder_ = std::move(builder);
1942 MdrunnerBuilder::MdrunnerBuilder(std::unique_ptr<MDModules> mdModules,
1943 compat::not_null<SimulationContext*> context) :
1944 impl_ {std::make_unique<Mdrunner::BuilderImplementation>(std::move(mdModules), context)}
1948 MdrunnerBuilder::~MdrunnerBuilder() = default;
1950 MdrunnerBuilder &MdrunnerBuilder::addSimulationMethod(const MdrunOptions &options,
1951 real forceWarningThreshold,
1952 const StartingBehavior startingBehavior)
1954 impl_->setExtraMdrunOptions(options, forceWarningThreshold, startingBehavior);
1958 MdrunnerBuilder &MdrunnerBuilder::addDomainDecomposition(const DomdecOptions &options)
1960 impl_->addDomdec(options);
1964 MdrunnerBuilder &MdrunnerBuilder::addNeighborList(int nstlist)
1966 impl_->addVerletList(nstlist);
1970 MdrunnerBuilder &MdrunnerBuilder::addReplicaExchange(const ReplicaExchangeParameters ¶ms)
1972 impl_->addReplicaExchange(params);
1976 MdrunnerBuilder &MdrunnerBuilder::addMultiSim(gmx_multisim_t* multisim)
1978 impl_->addMultiSim(multisim);
1982 MdrunnerBuilder &MdrunnerBuilder::addNonBonded(const char* nbpu_opt)
1984 impl_->addNonBonded(nbpu_opt);
1988 MdrunnerBuilder &MdrunnerBuilder::addElectrostatics(const char* pme_opt,
1989 const char* pme_fft_opt)
1991 // The builder method may become more general in the future, but in this version,
1992 // parameters for PME electrostatics are both required and the only parameters
1994 if (pme_opt && pme_fft_opt)
1996 impl_->addPME(pme_opt, pme_fft_opt);
2000 GMX_THROW(gmx::InvalidInputError("addElectrostatics() arguments must be non-null pointers."));
2005 MdrunnerBuilder &MdrunnerBuilder::addBondedTaskAssignment(const char* bonded_opt)
2007 impl_->addBondedTaskAssignment(bonded_opt);
2011 Mdrunner MdrunnerBuilder::build()
2013 return impl_->build();
2016 MdrunnerBuilder &MdrunnerBuilder::addHardwareOptions(const gmx_hw_opt_t &hardwareOptions)
2018 impl_->addHardwareOptions(hardwareOptions);
2022 MdrunnerBuilder &MdrunnerBuilder::addFilenames(ArrayRef<const t_filenm> filenames)
2024 impl_->addFilenames(filenames);
2028 MdrunnerBuilder &MdrunnerBuilder::addOutputEnvironment(gmx_output_env_t* outputEnvironment)
2030 impl_->addOutputEnvironment(outputEnvironment);
2034 MdrunnerBuilder &MdrunnerBuilder::addLogFile(t_fileio *logFileHandle)
2036 impl_->addLogFile(logFileHandle);
2040 MdrunnerBuilder &MdrunnerBuilder::addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder)
2042 impl_->addStopHandlerBuilder(std::move(builder));
2046 MdrunnerBuilder::MdrunnerBuilder(MdrunnerBuilder &&) noexcept = default;
2048 MdrunnerBuilder &MdrunnerBuilder::operator=(MdrunnerBuilder &&) noexcept = default;