2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
5 * Copyright (c) 2001-2004, The GROMACS development team.
6 * Copyright (c) 2011-2019,2020,2021, by the GROMACS development team, led by
7 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
8 * and including many others, as listed in the AUTHORS file in the
9 * top-level source directory and at http://www.gromacs.org.
11 * GROMACS is free software; you can redistribute it and/or
12 * modify it under the terms of the GNU Lesser General Public License
13 * as published by the Free Software Foundation; either version 2.1
14 * of the License, or (at your option) any later version.
16 * GROMACS is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
19 * Lesser General Public License for more details.
21 * You should have received a copy of the GNU Lesser General Public
22 * License along with GROMACS; if not, see
23 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
24 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
26 * If you want to redistribute modifications to GROMACS, please
27 * consider that scientific software is very special. Version
28 * control is crucial - bugs must be traceable. We will be happy to
29 * consider code for inclusion in the official distribution, but
30 * derived work must not be called official GROMACS. Details are found
31 * in the README & COPYING files - if they are missing, get the
32 * official version at http://www.gromacs.org.
34 * To help us fund GROMACS development, we humbly ask that you cite
35 * the research papers on the package. Check out http://www.gromacs.org.
39 * \brief Implements the MD runner routine calling all integrators.
41 * \author David van der Spoel <david.vanderspoel@icm.uu.se>
42 * \ingroup module_mdrun
59 #include "gromacs/commandline/filenm.h"
60 #include "gromacs/domdec/builder.h"
61 #include "gromacs/domdec/domdec.h"
62 #include "gromacs/domdec/domdec_struct.h"
63 #include "gromacs/domdec/gpuhaloexchange.h"
64 #include "gromacs/domdec/localatomsetmanager.h"
65 #include "gromacs/domdec/partition.h"
66 #include "gromacs/ewald/ewald_utils.h"
67 #include "gromacs/ewald/pme.h"
68 #include "gromacs/ewald/pme_gpu_program.h"
69 #include "gromacs/ewald/pme_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/gpu_utils.h"
77 #include "gromacs/hardware/cpuinfo.h"
78 #include "gromacs/hardware/detecthardware.h"
79 #include "gromacs/hardware/printhardware.h"
80 #include "gromacs/imd/imd.h"
81 #include "gromacs/listed_forces/disre.h"
82 #include "gromacs/listed_forces/gpubonded.h"
83 #include "gromacs/listed_forces/orires.h"
84 #include "gromacs/math/functions.h"
85 #include "gromacs/math/utilities.h"
86 #include "gromacs/math/vec.h"
87 #include "gromacs/mdlib/boxdeformation.h"
88 #include "gromacs/mdlib/broadcaststructs.h"
89 #include "gromacs/mdlib/calc_verletbuf.h"
90 #include "gromacs/mdlib/dispersioncorrection.h"
91 #include "gromacs/mdlib/enerdata_utils.h"
92 #include "gromacs/mdlib/force.h"
93 #include "gromacs/mdlib/forcerec.h"
94 #include "gromacs/mdlib/gmx_omp_nthreads.h"
95 #include "gromacs/mdlib/makeconstraints.h"
96 #include "gromacs/mdlib/md_support.h"
97 #include "gromacs/mdlib/mdatoms.h"
98 #include "gromacs/mdlib/membed.h"
99 #include "gromacs/mdlib/qmmm.h"
100 #include "gromacs/mdlib/sighandler.h"
101 #include "gromacs/mdlib/stophandler.h"
102 #include "gromacs/mdlib/updategroups.h"
103 #include "gromacs/mdrun/mdmodules.h"
104 #include "gromacs/mdrun/simulationcontext.h"
105 #include "gromacs/mdrunutility/handlerestart.h"
106 #include "gromacs/mdrunutility/logging.h"
107 #include "gromacs/mdrunutility/multisim.h"
108 #include "gromacs/mdrunutility/printtime.h"
109 #include "gromacs/mdrunutility/threadaffinity.h"
110 #include "gromacs/mdtypes/commrec.h"
111 #include "gromacs/mdtypes/enerdata.h"
112 #include "gromacs/mdtypes/fcdata.h"
113 #include "gromacs/mdtypes/group.h"
114 #include "gromacs/mdtypes/inputrec.h"
115 #include "gromacs/mdtypes/md_enums.h"
116 #include "gromacs/mdtypes/mdrunoptions.h"
117 #include "gromacs/mdtypes/observableshistory.h"
118 #include "gromacs/mdtypes/simulation_workload.h"
119 #include "gromacs/mdtypes/state.h"
120 #include "gromacs/mdtypes/state_propagator_data_gpu.h"
121 #include "gromacs/nbnxm/gpu_data_mgmt.h"
122 #include "gromacs/nbnxm/nbnxm.h"
123 #include "gromacs/nbnxm/pairlist_tuning.h"
124 #include "gromacs/pbcutil/pbc.h"
125 #include "gromacs/pulling/output.h"
126 #include "gromacs/pulling/pull.h"
127 #include "gromacs/pulling/pull_rotation.h"
128 #include "gromacs/restraint/manager.h"
129 #include "gromacs/restraint/restraintmdmodule.h"
130 #include "gromacs/restraint/restraintpotential.h"
131 #include "gromacs/swap/swapcoords.h"
132 #include "gromacs/taskassignment/decidegpuusage.h"
133 #include "gromacs/taskassignment/decidesimulationworkload.h"
134 #include "gromacs/taskassignment/resourcedivision.h"
135 #include "gromacs/taskassignment/taskassignment.h"
136 #include "gromacs/taskassignment/usergpuids.h"
137 #include "gromacs/timing/gpu_timing.h"
138 #include "gromacs/timing/wallcycle.h"
139 #include "gromacs/timing/wallcyclereporting.h"
140 #include "gromacs/topology/mtop_util.h"
141 #include "gromacs/trajectory/trajectoryframe.h"
142 #include "gromacs/utility/basenetwork.h"
143 #include "gromacs/utility/cstringutil.h"
144 #include "gromacs/utility/exceptions.h"
145 #include "gromacs/utility/fatalerror.h"
146 #include "gromacs/utility/filestream.h"
147 #include "gromacs/utility/gmxassert.h"
148 #include "gromacs/utility/gmxmpi.h"
149 #include "gromacs/utility/keyvaluetree.h"
150 #include "gromacs/utility/logger.h"
151 #include "gromacs/utility/loggerbuilder.h"
152 #include "gromacs/utility/mdmodulenotification.h"
153 #include "gromacs/utility/physicalnodecommunicator.h"
154 #include "gromacs/utility/pleasecite.h"
155 #include "gromacs/utility/programcontext.h"
156 #include "gromacs/utility/smalloc.h"
157 #include "gromacs/utility/stringutil.h"
159 #include "isimulator.h"
160 #include "replicaexchange.h"
161 #include "simulatorbuilder.h"
167 /*! \brief Manage any development feature flag variables encountered
169 * The use of dev features indicated by environment variables is
170 * logged in order to ensure that runs with such features enabled can
171 * be identified from their log and standard output. Any cross
172 * dependencies are also checked, and if unsatisfied, a fatal error
175 * Note that some development features overrides are applied already here:
176 * the GPU communication flags are set to false in non-tMPI and non-CUDA builds.
178 * \param[in] mdlog Logger object.
179 * \param[in] useGpuForNonbonded True if the nonbonded task is offloaded in this run.
180 * \param[in] pmeRunMode The PME run mode for this run
181 * \returns The object populated with development feature flags.
183 static DevelopmentFeatureFlags manageDevelopmentFeatures(const gmx::MDLogger& mdlog,
184 const bool useGpuForNonbonded,
185 const PmeRunMode pmeRunMode)
187 DevelopmentFeatureFlags devFlags;
189 // Some builds of GCC 5 give false positive warnings that these
190 // getenv results are ignored when clearly they are used.
191 #pragma GCC diagnostic push
192 #pragma GCC diagnostic ignored "-Wunused-result"
193 devFlags.enableGpuBufferOps = (getenv("GMX_USE_GPU_BUFFER_OPS") != nullptr)
194 && (GMX_GPU == GMX_GPU_CUDA) && useGpuForNonbonded;
195 devFlags.forceGpuUpdateDefault = (getenv("GMX_FORCE_UPDATE_DEFAULT_GPU") != nullptr) || GMX_FAHCORE;
196 devFlags.enableGpuHaloExchange =
197 (getenv("GMX_GPU_DD_COMMS") != nullptr && GMX_THREAD_MPI && (GMX_GPU == GMX_GPU_CUDA));
198 devFlags.enableGpuPmePPComm =
199 (getenv("GMX_GPU_PME_PP_COMMS") != nullptr && GMX_THREAD_MPI && (GMX_GPU == GMX_GPU_CUDA));
200 #pragma GCC diagnostic pop
202 if (devFlags.enableGpuBufferOps)
204 GMX_LOG(mdlog.warning)
206 .appendTextFormatted(
207 "This run uses the 'GPU buffer ops' feature, enabled by the "
208 "GMX_USE_GPU_BUFFER_OPS environment variable.");
211 if (devFlags.forceGpuUpdateDefault)
213 GMX_LOG(mdlog.warning)
215 .appendTextFormatted(
216 "This run will default to '-update gpu' as requested by the "
217 "GMX_FORCE_UPDATE_DEFAULT_GPU environment variable. GPU update with domain "
218 "decomposition lacks substantial testing and should be used with caution.");
221 if (devFlags.enableGpuHaloExchange)
223 if (useGpuForNonbonded)
225 if (!devFlags.enableGpuBufferOps)
227 GMX_LOG(mdlog.warning)
229 .appendTextFormatted(
230 "Enabling GPU buffer operations required by GMX_GPU_DD_COMMS "
231 "(equivalent with GMX_USE_GPU_BUFFER_OPS=1).");
232 devFlags.enableGpuBufferOps = true;
234 GMX_LOG(mdlog.warning)
236 .appendTextFormatted(
237 "This run uses the 'GPU halo exchange' feature, enabled by the "
238 "GMX_GPU_DD_COMMS environment variable.");
242 GMX_LOG(mdlog.warning)
244 .appendTextFormatted(
245 "GMX_GPU_DD_COMMS environment variable detected, but the 'GPU "
246 "halo exchange' feature will not be enabled as nonbonded interactions "
247 "are not offloaded.");
248 devFlags.enableGpuHaloExchange = false;
252 if (devFlags.enableGpuPmePPComm)
254 if (pmeRunMode == PmeRunMode::GPU)
256 GMX_LOG(mdlog.warning)
258 .appendTextFormatted(
259 "This run uses the 'GPU PME-PP communications' feature, enabled "
260 "by the GMX_GPU_PME_PP_COMMS environment variable.");
264 std::string clarification;
265 if (pmeRunMode == PmeRunMode::Mixed)
268 "PME FFT and gather are not offloaded to the GPU (PME is running in mixed "
273 clarification = "PME is not offloaded to the GPU.";
275 GMX_LOG(mdlog.warning)
278 "GMX_GPU_PME_PP_COMMS environment variable detected, but the "
279 "'GPU PME-PP communications' feature was not enabled as "
281 devFlags.enableGpuPmePPComm = false;
288 /*! \brief Barrier for safe simultaneous thread access to mdrunner data
290 * Used to ensure that the master thread does not modify mdrunner during copy
291 * on the spawned threads. */
292 static void threadMpiMdrunnerAccessBarrier()
295 MPI_Barrier(MPI_COMM_WORLD);
299 Mdrunner Mdrunner::cloneOnSpawnedThread() const
301 auto newRunner = Mdrunner(std::make_unique<MDModules>());
303 // All runners in the same process share a restraint manager resource because it is
304 // part of the interface to the client code, which is associated only with the
305 // original thread. Handles to the same resources can be obtained by copy.
307 newRunner.restraintManager_ = std::make_unique<RestraintManager>(*restraintManager_);
310 // Copy members of master runner.
311 // \todo Replace with builder when Simulation context and/or runner phases are better defined.
312 // Ref https://redmine.gromacs.org/issues/2587 and https://redmine.gromacs.org/issues/2375
313 newRunner.hw_opt = hw_opt;
314 newRunner.filenames = filenames;
316 newRunner.oenv = oenv;
317 newRunner.mdrunOptions = mdrunOptions;
318 newRunner.domdecOptions = domdecOptions;
319 newRunner.nbpu_opt = nbpu_opt;
320 newRunner.pme_opt = pme_opt;
321 newRunner.pme_fft_opt = pme_fft_opt;
322 newRunner.bonded_opt = bonded_opt;
323 newRunner.update_opt = update_opt;
324 newRunner.nstlist_cmdline = nstlist_cmdline;
325 newRunner.replExParams = replExParams;
326 newRunner.pforce = pforce;
327 // Give the spawned thread the newly created valid communicator
328 // for the simulation.
329 newRunner.communicator = MPI_COMM_WORLD;
331 newRunner.startingBehavior = startingBehavior;
332 newRunner.stopHandlerBuilder_ = std::make_unique<StopHandlerBuilder>(*stopHandlerBuilder_);
334 threadMpiMdrunnerAccessBarrier();
339 /*! \brief The callback used for running on spawned threads.
341 * Obtains the pointer to the master mdrunner object from the one
342 * argument permitted to the thread-launch API call, copies it to make
343 * a new runner for this thread, reinitializes necessary data, and
344 * proceeds to the simulation. */
345 static void mdrunner_start_fn(const void* arg)
349 auto masterMdrunner = reinterpret_cast<const gmx::Mdrunner*>(arg);
350 /* copy the arg list to make sure that it's thread-local. This
351 doesn't copy pointed-to items, of course; fnm, cr and fplog
352 are reset in the call below, all others should be const. */
353 gmx::Mdrunner mdrunner = masterMdrunner->cloneOnSpawnedThread();
356 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
360 void Mdrunner::spawnThreads(int numThreadsToLaunch)
363 /* now spawn new threads that start mdrunner_start_fn(), while
364 the main thread returns. Thread affinity is handled later. */
365 if (tMPI_Init_fn(TRUE, numThreadsToLaunch, TMPI_AFFINITY_NONE, mdrunner_start_fn,
366 static_cast<const void*>(this))
369 GMX_THROW(gmx::InternalError("Failed to spawn thread-MPI threads"));
372 // Give the master thread the newly created valid communicator for
374 communicator = MPI_COMM_WORLD;
375 threadMpiMdrunnerAccessBarrier();
377 GMX_UNUSED_VALUE(numThreadsToLaunch);
378 GMX_UNUSED_VALUE(mdrunner_start_fn);
384 /*! \brief Initialize variables for Verlet scheme simulation */
385 static void prepare_verlet_scheme(FILE* fplog,
389 const gmx_mtop_t* mtop,
391 bool makeGpuPairList,
392 const gmx::CpuInfo& cpuinfo)
394 /* For NVE simulations, we will retain the initial list buffer */
395 if (EI_DYNAMICS(ir->eI) && ir->verletbuf_tol > 0 && !(EI_MD(ir->eI) && ir->etc == etcNO))
397 /* Update the Verlet buffer size for the current run setup */
399 /* Here we assume SIMD-enabled kernels are being used. But as currently
400 * calc_verlet_buffer_size gives the same results for 4x8 and 4x4
401 * and 4x2 gives a larger buffer than 4x4, this is ok.
403 ListSetupType listType =
404 (makeGpuPairList ? ListSetupType::Gpu : ListSetupType::CpuSimdWhenSupported);
405 VerletbufListSetup listSetup = verletbufGetSafeListSetup(listType);
407 const real rlist_new =
408 calcVerletBufferSize(*mtop, det(box), *ir, ir->nstlist, ir->nstlist - 1, -1, listSetup);
410 if (rlist_new != ir->rlist)
412 if (fplog != nullptr)
415 "\nChanging rlist from %g to %g for non-bonded %dx%d atom kernels\n\n",
416 ir->rlist, rlist_new, listSetup.cluster_size_i, listSetup.cluster_size_j);
418 ir->rlist = rlist_new;
422 if (nstlist_cmdline > 0 && (!EI_DYNAMICS(ir->eI) || ir->verletbuf_tol <= 0))
424 gmx_fatal(FARGS, "Can not set nstlist without %s",
425 !EI_DYNAMICS(ir->eI) ? "dynamics" : "verlet-buffer-tolerance");
428 if (EI_DYNAMICS(ir->eI))
430 /* Set or try nstlist values */
431 increaseNstlist(fplog, cr, ir, nstlist_cmdline, mtop, box, makeGpuPairList, cpuinfo);
435 /*! \brief Override the nslist value in inputrec
437 * with value passed on the command line (if any)
439 static void override_nsteps_cmdline(const gmx::MDLogger& mdlog, int64_t nsteps_cmdline, t_inputrec* ir)
443 /* override with anything else than the default -2 */
444 if (nsteps_cmdline > -2)
446 char sbuf_steps[STEPSTRSIZE];
447 char sbuf_msg[STRLEN];
449 ir->nsteps = nsteps_cmdline;
450 if (EI_DYNAMICS(ir->eI) && nsteps_cmdline != -1)
453 "Overriding nsteps with value passed on the command line: %s steps, %.3g ps",
454 gmx_step_str(nsteps_cmdline, sbuf_steps), fabs(nsteps_cmdline * ir->delta_t));
458 sprintf(sbuf_msg, "Overriding nsteps with value passed on the command line: %s steps",
459 gmx_step_str(nsteps_cmdline, sbuf_steps));
462 GMX_LOG(mdlog.warning).asParagraph().appendText(sbuf_msg);
464 else if (nsteps_cmdline < -2)
466 gmx_fatal(FARGS, "Invalid nsteps value passed on the command line: %" PRId64, nsteps_cmdline);
468 /* Do nothing if nsteps_cmdline == -2 */
474 /*! \brief Return whether GPU acceleration of nonbondeds is supported with the given settings.
476 * If not, and if a warning may be issued, logs a warning about
477 * falling back to CPU code. With thread-MPI, only the first
478 * call to this function should have \c issueWarning true. */
479 static bool gpuAccelerationOfNonbondedIsUseful(const MDLogger& mdlog, const t_inputrec& ir, bool issueWarning)
481 bool gpuIsUseful = true;
484 if (ir.opts.ngener - ir.nwall > 1)
486 /* The GPU code does not support more than one energy group.
487 * If the user requested GPUs explicitly, a fatal error is given later.
491 "Multiple energy groups is not implemented for GPUs, falling back to the CPU. "
492 "For better performance, run on the GPU without energy groups and then do "
493 "gmx mdrun -rerun option on the trajectory with an energy group .tpr file.";
499 warning = "TPI is not implemented for GPUs.";
502 if (!gpuIsUseful && issueWarning)
504 GMX_LOG(mdlog.warning).asParagraph().appendText(warning);
510 //! Initializes the logger for mdrun.
511 static gmx::LoggerOwner buildLogger(FILE* fplog, const bool isSimulationMasterRank)
513 gmx::LoggerBuilder builder;
514 if (fplog != nullptr)
516 builder.addTargetFile(gmx::MDLogger::LogLevel::Info, fplog);
518 if (isSimulationMasterRank)
520 builder.addTargetStream(gmx::MDLogger::LogLevel::Warning, &gmx::TextOutputFile::standardError());
522 return builder.build();
525 //! Make a TaskTarget from an mdrun argument string.
526 static TaskTarget findTaskTarget(const char* optionString)
528 TaskTarget returnValue = TaskTarget::Auto;
530 if (strncmp(optionString, "auto", 3) == 0)
532 returnValue = TaskTarget::Auto;
534 else if (strncmp(optionString, "cpu", 3) == 0)
536 returnValue = TaskTarget::Cpu;
538 else if (strncmp(optionString, "gpu", 3) == 0)
540 returnValue = TaskTarget::Gpu;
544 GMX_ASSERT(false, "Option string should have been checked for sanity already");
550 //! Finish run, aggregate data to print performance info.
551 static void finish_run(FILE* fplog,
552 const gmx::MDLogger& mdlog,
554 const t_inputrec* inputrec,
556 gmx_wallcycle_t wcycle,
557 gmx_walltime_accounting_t walltime_accounting,
558 nonbonded_verlet_t* nbv,
559 const gmx_pme_t* pme,
563 double nbfs = 0, mflop = 0;
564 double elapsed_time, elapsed_time_over_all_ranks, elapsed_time_over_all_threads,
565 elapsed_time_over_all_threads_over_all_ranks;
566 /* Control whether it is valid to print a report. Only the
567 simulation master may print, but it should not do so if the run
568 terminated e.g. before a scheduled reset step. This is
569 complicated by the fact that PME ranks are unaware of the
570 reason why they were sent a pmerecvqxFINISH. To avoid
571 communication deadlocks, we always do the communication for the
572 report, even if we've decided not to write the report, because
573 how long it takes to finish the run is not important when we've
574 decided not to report on the simulation performance.
576 Further, we only report performance for dynamical integrators,
577 because those are the only ones for which we plan to
578 consider doing any optimizations. */
579 bool printReport = EI_DYNAMICS(inputrec->eI) && SIMMASTER(cr);
581 if (printReport && !walltime_accounting_get_valid_finish(walltime_accounting))
583 GMX_LOG(mdlog.warning)
585 .appendText("Simulation ended prematurely, no performance report will be written.");
590 std::unique_ptr<t_nrnb> nrnbTotalStorage;
593 nrnbTotalStorage = std::make_unique<t_nrnb>();
594 nrnb_tot = nrnbTotalStorage.get();
596 MPI_Allreduce(nrnb->n, nrnb_tot->n, eNRNB, MPI_DOUBLE, MPI_SUM, cr->mpi_comm_mysim);
604 elapsed_time = walltime_accounting_get_time_since_reset(walltime_accounting);
605 elapsed_time_over_all_threads =
606 walltime_accounting_get_time_since_reset_over_all_threads(walltime_accounting);
610 /* reduce elapsed_time over all MPI ranks in the current simulation */
611 MPI_Allreduce(&elapsed_time, &elapsed_time_over_all_ranks, 1, MPI_DOUBLE, MPI_SUM,
613 elapsed_time_over_all_ranks /= cr->nnodes;
614 /* Reduce elapsed_time_over_all_threads over all MPI ranks in the
615 * current simulation. */
616 MPI_Allreduce(&elapsed_time_over_all_threads, &elapsed_time_over_all_threads_over_all_ranks,
617 1, MPI_DOUBLE, MPI_SUM, cr->mpi_comm_mysim);
622 elapsed_time_over_all_ranks = elapsed_time;
623 elapsed_time_over_all_threads_over_all_ranks = elapsed_time_over_all_threads;
628 print_flop(fplog, nrnb_tot, &nbfs, &mflop);
631 if (thisRankHasDuty(cr, DUTY_PP) && DOMAINDECOMP(cr))
633 print_dd_statistics(cr, inputrec, fplog);
636 /* TODO Move the responsibility for any scaling by thread counts
637 * to the code that handled the thread region, so that there's a
638 * mechanism to keep cycle counting working during the transition
639 * to task parallelism. */
640 int nthreads_pp = gmx_omp_nthreads_get(emntNonbonded);
641 int nthreads_pme = gmx_omp_nthreads_get(emntPME);
642 wallcycle_scale_by_num_threads(wcycle, thisRankHasDuty(cr, DUTY_PME) && !thisRankHasDuty(cr, DUTY_PP),
643 nthreads_pp, nthreads_pme);
644 auto cycle_sum(wallcycle_sum(cr, wcycle));
648 auto nbnxn_gpu_timings =
649 (nbv != nullptr && nbv->useGpu()) ? Nbnxm::gpu_get_timings(nbv->gpu_nbv) : nullptr;
650 gmx_wallclock_gpu_pme_t pme_gpu_timings = {};
652 if (pme_gpu_task_enabled(pme))
654 pme_gpu_get_timings(pme, &pme_gpu_timings);
656 wallcycle_print(fplog, mdlog, cr->nnodes, cr->npmenodes, nthreads_pp, nthreads_pme,
657 elapsed_time_over_all_ranks, wcycle, cycle_sum, nbnxn_gpu_timings,
660 if (EI_DYNAMICS(inputrec->eI))
662 delta_t = inputrec->delta_t;
667 print_perf(fplog, elapsed_time_over_all_threads_over_all_ranks, elapsed_time_over_all_ranks,
668 walltime_accounting_get_nsteps_done_since_reset(walltime_accounting),
669 delta_t, nbfs, mflop);
673 print_perf(stderr, elapsed_time_over_all_threads_over_all_ranks, elapsed_time_over_all_ranks,
674 walltime_accounting_get_nsteps_done_since_reset(walltime_accounting),
675 delta_t, nbfs, mflop);
680 int Mdrunner::mdrunner()
683 t_forcerec* fr = nullptr;
684 t_fcdata* fcd = nullptr;
685 real ewaldcoeff_q = 0;
686 real ewaldcoeff_lj = 0;
687 int nChargePerturbed = -1, nTypePerturbed = 0;
688 gmx_wallcycle_t wcycle;
689 gmx_walltime_accounting_t walltime_accounting = nullptr;
690 gmx_membed_t* membed = nullptr;
691 gmx_hw_info_t* hwinfo = nullptr;
693 /* CAUTION: threads may be started later on in this function, so
694 cr doesn't reflect the final parallel state right now */
697 /* TODO: inputrec should tell us whether we use an algorithm, not a file option */
698 const bool doEssentialDynamics = opt2bSet("-ei", filenames.size(), filenames.data());
699 const bool doMembed = opt2bSet("-membed", filenames.size(), filenames.data());
700 const bool doRerun = mdrunOptions.rerun;
702 // Handle task-assignment related user options.
703 EmulateGpuNonbonded emulateGpuNonbonded =
704 (getenv("GMX_EMULATE_GPU") != nullptr ? EmulateGpuNonbonded::Yes : EmulateGpuNonbonded::No);
706 std::vector<int> userGpuTaskAssignment;
709 userGpuTaskAssignment = parseUserTaskAssignmentString(hw_opt.userGpuTaskAssignment);
711 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
712 auto nonbondedTarget = findTaskTarget(nbpu_opt);
713 auto pmeTarget = findTaskTarget(pme_opt);
714 auto pmeFftTarget = findTaskTarget(pme_fft_opt);
715 auto bondedTarget = findTaskTarget(bonded_opt);
716 auto updateTarget = findTaskTarget(update_opt);
718 FILE* fplog = nullptr;
719 // If we are appending, we don't write log output because we need
720 // to check that the old log file matches what the checkpoint file
721 // expects. Otherwise, we should start to write log output now if
722 // there is a file ready for it.
723 if (logFileHandle != nullptr && startingBehavior != StartingBehavior::RestartWithAppending)
725 fplog = gmx_fio_getfp(logFileHandle);
727 const bool isSimulationMasterRank = findIsSimulationMasterRank(ms, communicator);
728 gmx::LoggerOwner logOwner(buildLogger(fplog, isSimulationMasterRank));
729 gmx::MDLogger mdlog(logOwner.logger());
731 // TODO The thread-MPI master rank makes a working
732 // PhysicalNodeCommunicator here, but it gets rebuilt by all ranks
733 // after the threads have been launched. This works because no use
734 // is made of that communicator until after the execution paths
735 // have rejoined. But it is likely that we can improve the way
736 // this is expressed, e.g. by expressly running detection only the
737 // master rank for thread-MPI, rather than relying on the mutex
738 // and reference count.
739 PhysicalNodeCommunicator physicalNodeComm(communicator, gmx_physicalnode_id_hash());
740 hwinfo = gmx_detect_hardware(mdlog, physicalNodeComm);
742 gmx_print_detected_hardware(fplog, isSimulationMasterRank && isMasterSim(ms), mdlog, hwinfo);
744 std::vector<int> gpuIdsToUse = makeGpuIdsToUse(hwinfo->gpu_info, hw_opt.gpuIdsAvailable);
746 // Print citation requests after all software/hardware printing
747 pleaseCiteGromacs(fplog);
749 // TODO Replace this by unique_ptr once t_inputrec is C++
750 t_inputrec inputrecInstance;
751 t_inputrec* inputrec = nullptr;
752 std::unique_ptr<t_state> globalState;
754 auto partialDeserializedTpr = std::make_unique<PartialDeserializedTprFile>();
756 if (isSimulationMasterRank)
758 /* Only the master rank has the global state */
759 globalState = std::make_unique<t_state>();
761 /* Read (nearly) all data required for the simulation
762 * and keep the partly serialized tpr contents to send to other ranks later
764 *partialDeserializedTpr = read_tpx_state(ftp2fn(efTPR, filenames.size(), filenames.data()),
765 &inputrecInstance, globalState.get(), &mtop);
766 inputrec = &inputrecInstance;
769 /* Check and update the hardware options for internal consistency */
770 checkAndUpdateHardwareOptions(mdlog, &hw_opt, isSimulationMasterRank, domdecOptions.numPmeRanks,
773 if (GMX_THREAD_MPI && isSimulationMasterRank)
775 bool useGpuForNonbonded = false;
776 bool useGpuForPme = false;
779 GMX_RELEASE_ASSERT(inputrec != nullptr, "Keep the compiler happy");
781 // If the user specified the number of ranks, then we must
782 // respect that, but in default mode, we need to allow for
783 // the number of GPUs to choose the number of ranks.
784 auto canUseGpuForNonbonded = buildSupportsNonbondedOnGpu(nullptr);
785 useGpuForNonbonded = decideWhetherToUseGpusForNonbondedWithThreadMpi(
786 nonbondedTarget, gpuIdsToUse, userGpuTaskAssignment, emulateGpuNonbonded,
787 canUseGpuForNonbonded,
788 gpuAccelerationOfNonbondedIsUseful(mdlog, *inputrec, GMX_THREAD_MPI),
789 hw_opt.nthreads_tmpi);
790 useGpuForPme = decideWhetherToUseGpusForPmeWithThreadMpi(
791 useGpuForNonbonded, pmeTarget, gpuIdsToUse, userGpuTaskAssignment, *hwinfo,
792 *inputrec, mtop, hw_opt.nthreads_tmpi, domdecOptions.numPmeRanks);
794 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
796 /* Determine how many thread-MPI ranks to start.
798 * TODO Over-writing the user-supplied value here does
799 * prevent any possible subsequent checks from working
801 hw_opt.nthreads_tmpi = get_nthreads_mpi(hwinfo, &hw_opt, gpuIdsToUse, useGpuForNonbonded,
802 useGpuForPme, inputrec, &mtop, mdlog, doMembed);
804 // Now start the threads for thread MPI.
805 spawnThreads(hw_opt.nthreads_tmpi);
806 // The spawned threads enter mdrunner() and execution of
807 // master and spawned threads joins at the end of this block.
808 physicalNodeComm = PhysicalNodeCommunicator(communicator, gmx_physicalnode_id_hash());
811 GMX_RELEASE_ASSERT(communicator == MPI_COMM_WORLD, "Must have valid world communicator");
812 CommrecHandle crHandle = init_commrec(communicator, ms);
813 t_commrec* cr = crHandle.get();
814 GMX_RELEASE_ASSERT(cr != nullptr, "Must have valid commrec");
818 /* now broadcast everything to the non-master nodes/threads: */
819 if (!isSimulationMasterRank)
821 inputrec = &inputrecInstance;
823 init_parallel(cr, inputrec, &mtop, partialDeserializedTpr.get());
825 GMX_RELEASE_ASSERT(inputrec != nullptr, "All ranks should have a valid inputrec now");
826 partialDeserializedTpr.reset(nullptr);
828 // Now the number of ranks is known to all ranks, and each knows
829 // the inputrec read by the master rank. The ranks can now all run
830 // the task-deciding functions and will agree on the result
831 // without needing to communicate.
833 // TODO Should we do the communication in debug mode to support
834 // having an assertion?
835 const bool useDomainDecomposition = (PAR(cr) && !(EI_TPI(inputrec->eI) || inputrec->eI == eiNM));
837 // Note that these variables describe only their own node.
839 // Note that when bonded interactions run on a GPU they always run
840 // alongside a nonbonded task, so do not influence task assignment
841 // even though they affect the force calculation workload.
842 bool useGpuForNonbonded = false;
843 bool useGpuForPme = false;
844 bool useGpuForBonded = false;
845 bool useGpuForUpdate = false;
846 bool gpusWereDetected = hwinfo->ngpu_compatible_tot > 0;
849 // It's possible that there are different numbers of GPUs on
850 // different nodes, which is the user's responsibility to
851 // handle. If unsuitable, we will notice that during task
853 auto canUseGpuForNonbonded = buildSupportsNonbondedOnGpu(nullptr);
854 useGpuForNonbonded = decideWhetherToUseGpusForNonbonded(
855 nonbondedTarget, userGpuTaskAssignment, emulateGpuNonbonded, canUseGpuForNonbonded,
856 gpuAccelerationOfNonbondedIsUseful(mdlog, *inputrec, !GMX_THREAD_MPI), gpusWereDetected);
857 useGpuForPme = decideWhetherToUseGpusForPme(
858 useGpuForNonbonded, pmeTarget, userGpuTaskAssignment, *hwinfo, *inputrec, mtop,
859 cr->nnodes, domdecOptions.numPmeRanks, gpusWereDetected);
860 auto canUseGpuForBonded = buildSupportsGpuBondeds(nullptr)
861 && inputSupportsGpuBondeds(*inputrec, mtop, nullptr);
862 useGpuForBonded = decideWhetherToUseGpusForBonded(
863 useGpuForNonbonded, useGpuForPme, bondedTarget, canUseGpuForBonded,
864 EVDW_PME(inputrec->vdwtype), EEL_PME_EWALD(inputrec->coulombtype),
865 domdecOptions.numPmeRanks, gpusWereDetected);
867 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
869 const PmeRunMode pmeRunMode = determinePmeRunMode(useGpuForPme, pmeFftTarget, *inputrec);
871 // Initialize development feature flags that enabled by environment variable
872 // and report those features that are enabled.
873 const DevelopmentFeatureFlags devFlags =
874 manageDevelopmentFeatures(mdlog, useGpuForNonbonded, pmeRunMode);
876 const bool inputIsCompatibleWithModularSimulator = ModularSimulator::isInputCompatible(
877 false, inputrec, doRerun, mtop, ms, replExParams, nullptr, doEssentialDynamics, doMembed);
878 const bool useModularSimulator = inputIsCompatibleWithModularSimulator
879 && !(getenv("GMX_DISABLE_MODULAR_SIMULATOR") != nullptr);
882 // TODO: hide restraint implementation details from Mdrunner.
883 // There is nothing unique about restraints at this point as far as the
884 // Mdrunner is concerned. The Mdrunner should just be getting a sequence of
885 // factory functions from the SimulationContext on which to call mdModules_->add().
886 // TODO: capture all restraints into a single RestraintModule, passed to the runner builder.
887 for (auto&& restraint : restraintManager_->getRestraints())
889 auto module = RestraintMDModule::create(restraint, restraint->sites());
890 mdModules_->add(std::move(module));
893 // TODO: Error handling
894 mdModules_->assignOptionsToModules(*inputrec->params, nullptr);
895 const auto& mdModulesNotifier = mdModules_->notifier().notifier_;
897 if (inputrec->internalParameters != nullptr)
899 mdModulesNotifier.notify(*inputrec->internalParameters);
902 if (fplog != nullptr)
904 pr_inputrec(fplog, 0, "Input Parameters", inputrec, FALSE);
905 fprintf(fplog, "\n");
910 /* In rerun, set velocities to zero if present */
911 if (doRerun && ((globalState->flags & (1 << estV)) != 0))
913 // rerun does not use velocities
917 "Rerun trajectory contains velocities. Rerun does only evaluate "
918 "potential energy and forces. The velocities will be ignored.");
919 for (int i = 0; i < globalState->natoms; i++)
921 clear_rvec(globalState->v[i]);
923 globalState->flags &= ~(1 << estV);
926 /* now make sure the state is initialized and propagated */
927 set_state_entries(globalState.get(), inputrec, useModularSimulator);
930 /* NM and TPI parallelize over force/energy calculations, not atoms,
931 * so we need to initialize and broadcast the global state.
933 if (inputrec->eI == eiNM || inputrec->eI == eiTPI)
937 globalState = std::make_unique<t_state>();
939 broadcastStateWithoutDynamics(cr, globalState.get());
942 /* A parallel command line option consistency check that we can
943 only do after any threads have started. */
945 && (domdecOptions.numCells[XX] > 1 || domdecOptions.numCells[YY] > 1
946 || domdecOptions.numCells[ZZ] > 1 || domdecOptions.numPmeRanks > 0))
949 "The -dd or -npme option request a parallel simulation, "
951 "but %s was compiled without threads or MPI enabled",
952 output_env_get_program_display_name(oenv));
954 "but the number of MPI-threads (option -ntmpi) is not set or is 1");
956 "but %s was not started through mpirun/mpiexec or only one rank was requested "
957 "through mpirun/mpiexec",
958 output_env_get_program_display_name(oenv));
962 if (doRerun && (EI_ENERGY_MINIMIZATION(inputrec->eI) || eiNM == inputrec->eI))
965 "The .mdp file specified an energy mininization or normal mode algorithm, and "
966 "these are not compatible with mdrun -rerun");
969 if (!(EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype)))
971 if (domdecOptions.numPmeRanks > 0)
973 gmx_fatal_collective(FARGS, cr->mpi_comm_mysim, MASTER(cr),
974 "PME-only ranks are requested, but the system does not use PME "
975 "for electrostatics or LJ");
978 domdecOptions.numPmeRanks = 0;
981 if (useGpuForNonbonded && domdecOptions.numPmeRanks < 0)
983 /* With NB GPUs we don't automatically use PME-only CPU ranks. PME ranks can
984 * improve performance with many threads per GPU, since our OpenMP
985 * scaling is bad, but it's difficult to automate the setup.
987 domdecOptions.numPmeRanks = 0;
991 if (domdecOptions.numPmeRanks < 0)
993 domdecOptions.numPmeRanks = 0;
994 // TODO possibly print a note that one can opt-in for a separate PME GPU rank?
998 GMX_RELEASE_ASSERT(domdecOptions.numPmeRanks <= 1,
999 "PME GPU decomposition is not supported");
1003 /* NMR restraints must be initialized before load_checkpoint,
1004 * since with time averaging the history is added to t_state.
1005 * For proper consistency check we therefore need to extend
1007 * So the PME-only nodes (if present) will also initialize
1008 * the distance restraints.
1012 /* This needs to be called before read_checkpoint to extend the state */
1013 init_disres(fplog, &mtop, inputrec, cr, ms, fcd, globalState.get(), replExParams.exchangeInterval > 0);
1015 init_orires(fplog, &mtop, inputrec, cr, ms, globalState.get(), &(fcd->orires));
1017 auto deform = prepareBoxDeformation(globalState->box, cr, *inputrec);
1020 /* We have to remember the generation's first step before reading checkpoint.
1021 This way, we can report to the F@H core both the generation's first step
1022 and the restored first step, thus making it able to distinguish between
1023 an interruption/resume and start of the n-th generation simulation.
1024 Having this information, the F@H core can correctly calculate and report
1027 int gen_first_step = 0;
1030 gen_first_step = inputrec->init_step;
1034 ObservablesHistory observablesHistory = {};
1036 if (startingBehavior != StartingBehavior::NewSimulation)
1038 /* Check if checkpoint file exists before doing continuation.
1039 * This way we can use identical input options for the first and subsequent runs...
1041 if (mdrunOptions.numStepsCommandline > -2)
1043 /* Temporarily set the number of steps to unlimited to avoid
1044 * triggering the nsteps check in load_checkpoint().
1045 * This hack will go away soon when the -nsteps option is removed.
1047 inputrec->nsteps = -1;
1050 load_checkpoint(opt2fn_master("-cpi", filenames.size(), filenames.data(), cr),
1051 logFileHandle, cr, domdecOptions.numCells, inputrec, globalState.get(),
1052 &observablesHistory, mdrunOptions.reproducible, mdModules_->notifier());
1054 if (startingBehavior == StartingBehavior::RestartWithAppending && logFileHandle)
1056 // Now we can start normal logging to the truncated log file.
1057 fplog = gmx_fio_getfp(logFileHandle);
1058 prepareLogAppending(fplog);
1059 logOwner = buildLogger(fplog, MASTER(cr));
1060 mdlog = logOwner.logger();
1067 fcRegisterSteps(inputrec->nsteps + inputrec->init_step, gen_first_step);
1071 if (mdrunOptions.numStepsCommandline > -2)
1076 "The -nsteps functionality is deprecated, and may be removed in a future "
1078 "Consider using gmx convert-tpr -nsteps or changing the appropriate .mdp "
1081 /* override nsteps with value set on the commandline */
1082 override_nsteps_cmdline(mdlog, mdrunOptions.numStepsCommandline, inputrec);
1086 copy_mat(globalState->box, box);
1091 gmx_bcast(sizeof(box), box, cr);
1094 if (inputrec->cutoff_scheme != ecutsVERLET)
1097 "This group-scheme .tpr file can no longer be run by mdrun. Please update to the "
1098 "Verlet scheme, or use an earlier version of GROMACS if necessary.");
1100 /* Update rlist and nstlist. */
1101 prepare_verlet_scheme(fplog, cr, inputrec, nstlist_cmdline, &mtop, box,
1102 useGpuForNonbonded || (emulateGpuNonbonded == EmulateGpuNonbonded::Yes),
1105 const bool prefer1DAnd1PulseDD = (devFlags.enableGpuHaloExchange && useGpuForNonbonded);
1106 // This builder is necessary while we have multi-part construction
1107 // of DD. Before DD is constructed, we use the existence of
1108 // the builder object to indicate that further construction of DD
1110 std::unique_ptr<DomainDecompositionBuilder> ddBuilder;
1111 if (useDomainDecomposition)
1113 ddBuilder = std::make_unique<DomainDecompositionBuilder>(
1114 mdlog, cr, domdecOptions, mdrunOptions, prefer1DAnd1PulseDD, mtop, *inputrec, box,
1115 positionsFromStatePointer(globalState.get()));
1119 /* PME, if used, is done on all nodes with 1D decomposition */
1121 cr->duty = (DUTY_PP | DUTY_PME);
1123 if (inputrec->ePBC == epbcSCREW)
1125 gmx_fatal(FARGS, "pbc=screw is only implemented with domain decomposition");
1129 // Produce the task assignment for this rank.
1130 GpuTaskAssignmentsBuilder gpuTaskAssignmentsBuilder;
1131 GpuTaskAssignments gpuTaskAssignments = gpuTaskAssignmentsBuilder.build(
1132 gpuIdsToUse, userGpuTaskAssignment, *hwinfo, communicator, physicalNodeComm,
1133 nonbondedTarget, pmeTarget, bondedTarget, updateTarget, useGpuForNonbonded,
1134 useGpuForPme, thisRankHasDuty(cr, DUTY_PP),
1135 // TODO cr->duty & DUTY_PME should imply that a PME
1136 // algorithm is active, but currently does not.
1137 EEL_PME(inputrec->coulombtype) && thisRankHasDuty(cr, DUTY_PME));
1139 // Get the device handles for the modules, nullptr when no task is assigned.
1140 gmx_device_info_t* nonbondedDeviceInfo = gpuTaskAssignments.initNonbondedDevice(cr);
1141 gmx_device_info_t* pmeDeviceInfo = gpuTaskAssignments.initPmeDevice();
1143 // TODO Initialize GPU streams here.
1145 // TODO Currently this is always built, yet DD partition code
1146 // checks if it is built before using it. Probably it should
1147 // become an MDModule that is made only when another module
1148 // requires it (e.g. pull, CompEl, density fitting), so that we
1149 // don't update the local atom sets unilaterally every step.
1150 LocalAtomSetManager atomSets;
1153 // TODO Pass the GPU streams to ddBuilder to use in buffer
1154 // transfers (e.g. halo exchange)
1155 cr->dd = ddBuilder->build(&atomSets);
1156 // The builder's job is done, so destruct it
1157 ddBuilder.reset(nullptr);
1158 // Note that local state still does not exist yet.
1161 // The GPU update is decided here because we need to know whether the constraints or
1162 // SETTLEs can span accross the domain borders (i.e. whether or not update groups are
1163 // defined). This is only known after DD is initialized, hence decision on using GPU
1164 // update is done so late.
1167 const bool useUpdateGroups = cr->dd ? ddUsesUpdateGroups(*cr->dd) : false;
1169 useGpuForUpdate = decideWhetherToUseGpuForUpdate(
1170 useDomainDecomposition, useUpdateGroups, pmeRunMode, domdecOptions.numPmeRanks > 0,
1171 useGpuForNonbonded, updateTarget, gpusWereDetected, *inputrec, mtop,
1172 doEssentialDynamics, gmx_mtop_ftype_count(mtop, F_ORIRES) > 0,
1173 replExParams.exchangeInterval > 0, doRerun, devFlags, mdlog);
1175 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1177 const bool printHostName = (cr->nnodes > 1);
1178 gpuTaskAssignments.reportGpuUsage(mdlog, printHostName, useGpuForBonded, pmeRunMode, useGpuForUpdate);
1180 // If the user chose a task assignment, give them some hints
1181 // where appropriate.
1182 if (!userGpuTaskAssignment.empty())
1184 gpuTaskAssignments.logPerformanceHints(mdlog, ssize(gpuIdsToUse));
1189 /* After possible communicator splitting in make_dd_communicators.
1190 * we can set up the intra/inter node communication.
1192 gmx_setup_nodecomm(fplog, cr);
1198 GMX_LOG(mdlog.warning)
1200 .appendTextFormatted(
1201 "This is simulation %d out of %d running as a composite GROMACS\n"
1202 "multi-simulation job. Setup for this simulation:\n",
1205 GMX_LOG(mdlog.warning)
1206 .appendTextFormatted("Using %d MPI %s\n", cr->nnodes,
1208 cr->nnodes == 1 ? "thread" : "threads"
1210 cr->nnodes == 1 ? "process" : "processes"
1216 // If mdrun -pin auto honors any affinity setting that already
1217 // exists. If so, it is nice to provide feedback about whether
1218 // that existing affinity setting was from OpenMP or something
1219 // else, so we run this code both before and after we initialize
1220 // the OpenMP support.
1221 gmx_check_thread_affinity_set(mdlog, &hw_opt, hwinfo->nthreads_hw_avail, FALSE);
1222 /* Check and update the number of OpenMP threads requested */
1223 checkAndUpdateRequestedNumOpenmpThreads(&hw_opt, *hwinfo, cr, ms, physicalNodeComm.size_,
1224 pmeRunMode, mtop, *inputrec);
1226 gmx_omp_nthreads_init(mdlog, cr, hwinfo->nthreads_hw_avail, physicalNodeComm.size_,
1227 hw_opt.nthreads_omp, hw_opt.nthreads_omp_pme, !thisRankHasDuty(cr, DUTY_PP));
1229 // Enable FP exception detection, but not in
1230 // Release mode and not for compilers with known buggy FP
1231 // exception support (clang with any optimization) or suspected
1232 // buggy FP exception support (gcc 7.* with optimization).
1233 #if !defined NDEBUG \
1234 && !((defined __clang__ || (defined(__GNUC__) && !defined(__ICC) && __GNUC__ == 7)) \
1235 && defined __OPTIMIZE__)
1236 const bool bEnableFPE = true;
1238 const bool bEnableFPE = false;
1240 // FIXME - reconcile with gmx_feenableexcept() call from CommandLineModuleManager::run()
1243 gmx_feenableexcept();
1246 /* Now that we know the setup is consistent, check for efficiency */
1247 check_resource_division_efficiency(hwinfo, gpuTaskAssignments.thisRankHasAnyGpuTask(),
1248 mdrunOptions.ntompOptionIsSet, cr, mdlog);
1250 /* getting number of PP/PME threads on this MPI / tMPI rank.
1251 PME: env variable should be read only on one node to make sure it is
1252 identical everywhere;
1254 const int numThreadsOnThisRank = thisRankHasDuty(cr, DUTY_PP) ? gmx_omp_nthreads_get(emntNonbonded)
1255 : gmx_omp_nthreads_get(emntPME);
1256 checkHardwareOversubscription(numThreadsOnThisRank, cr->nodeid, *hwinfo->hardwareTopology,
1257 physicalNodeComm, mdlog);
1259 // Enable Peer access between GPUs where available
1260 // Only for DD, only master PP rank needs to perform setup, and only if thread MPI plus
1261 // any of the GPU communication features are active.
1262 if (DOMAINDECOMP(cr) && MASTER(cr) && thisRankHasDuty(cr, DUTY_PP) && GMX_THREAD_MPI
1263 && (devFlags.enableGpuHaloExchange || devFlags.enableGpuPmePPComm))
1265 setupGpuDevicePeerAccess(gpuIdsToUse, mdlog);
1268 if (hw_opt.threadAffinity != ThreadAffinity::Off)
1270 /* Before setting affinity, check whether the affinity has changed
1271 * - which indicates that probably the OpenMP library has changed it
1272 * since we first checked).
1274 gmx_check_thread_affinity_set(mdlog, &hw_opt, hwinfo->nthreads_hw_avail, TRUE);
1276 int numThreadsOnThisNode, intraNodeThreadOffset;
1277 analyzeThreadsOnThisNode(physicalNodeComm, numThreadsOnThisRank, &numThreadsOnThisNode,
1278 &intraNodeThreadOffset);
1280 /* Set the CPU affinity */
1281 gmx_set_thread_affinity(mdlog, cr, &hw_opt, *hwinfo->hardwareTopology, numThreadsOnThisRank,
1282 numThreadsOnThisNode, intraNodeThreadOffset, nullptr);
1285 if (mdrunOptions.timingOptions.resetStep > -1)
1290 "The -resetstep functionality is deprecated, and may be removed in a "
1293 wcycle = wallcycle_init(fplog, mdrunOptions.timingOptions.resetStep, cr);
1297 /* Master synchronizes its value of reset_counters with all nodes
1298 * including PME only nodes */
1299 int64_t reset_counters = wcycle_get_reset_counters(wcycle);
1300 gmx_bcast_sim(sizeof(reset_counters), &reset_counters, cr);
1301 wcycle_set_reset_counters(wcycle, reset_counters);
1304 // Membrane embedding must be initialized before we call init_forcerec()
1309 fprintf(stderr, "Initializing membed");
1311 /* Note that membed cannot work in parallel because mtop is
1312 * changed here. Fix this if we ever want to make it run with
1313 * multiple ranks. */
1314 membed = init_membed(fplog, filenames.size(), filenames.data(), &mtop, inputrec,
1315 globalState.get(), cr, &mdrunOptions.checkpointOptions.period);
1318 const bool thisRankHasPmeGpuTask = gpuTaskAssignments.thisRankHasPmeGpuTask();
1319 std::unique_ptr<MDAtoms> mdAtoms;
1320 std::unique_ptr<gmx_vsite_t> vsite;
1323 if (thisRankHasDuty(cr, DUTY_PP))
1325 mdModulesNotifier.notify(*cr);
1326 mdModulesNotifier.notify(&atomSets);
1327 mdModulesNotifier.notify(PeriodicBoundaryConditionType{ inputrec->ePBC });
1328 mdModulesNotifier.notify(SimulationTimeStep{ inputrec->delta_t });
1329 /* Initiate forcerecord */
1330 fr = new t_forcerec;
1331 fr->forceProviders = mdModules_->initForceProviders();
1332 init_forcerec(fplog, mdlog, fr, fcd, inputrec, &mtop, cr, box,
1333 opt2fn("-table", filenames.size(), filenames.data()),
1334 opt2fn("-tablep", filenames.size(), filenames.data()),
1335 opt2fns("-tableb", filenames.size(), filenames.data()), *hwinfo,
1336 nonbondedDeviceInfo, useGpuForBonded,
1337 pmeRunMode == PmeRunMode::GPU && !thisRankHasDuty(cr, DUTY_PME), pforce, wcycle);
1339 // TODO Move this to happen during domain decomposition setup,
1340 // once stream and event handling works well with that.
1341 // TODO remove need to pass local stream into GPU halo exchange - Redmine #3093
1342 if (havePPDomainDecomposition(cr) && prefer1DAnd1PulseDD && is1DAnd1PulseDD(*cr->dd))
1344 GMX_RELEASE_ASSERT(devFlags.enableGpuBufferOps,
1345 "Must use GMX_USE_GPU_BUFFER_OPS=1 to use GMX_GPU_DD_COMMS=1");
1347 Nbnxm::gpu_get_command_stream(fr->nbv->gpu_nbv, InteractionLocality::Local);
1348 void* streamNonLocal =
1349 Nbnxm::gpu_get_command_stream(fr->nbv->gpu_nbv, InteractionLocality::NonLocal);
1350 GMX_LOG(mdlog.warning)
1352 .appendTextFormatted(
1353 "NOTE: This run uses the 'GPU halo exchange' feature, enabled by the "
1354 "GMX_GPU_DD_COMMS environment variable.");
1355 cr->dd->gpuHaloExchange = std::make_unique<GpuHaloExchange>(
1356 cr->dd, cr->mpi_comm_mysim, streamLocal, streamNonLocal);
1359 /* Initialize the mdAtoms structure.
1360 * mdAtoms is not filled with atom data,
1361 * as this can not be done now with domain decomposition.
1363 mdAtoms = makeMDAtoms(fplog, mtop, *inputrec, thisRankHasPmeGpuTask);
1364 if (globalState && thisRankHasPmeGpuTask)
1366 // The pinning of coordinates in the global state object works, because we only use
1367 // PME on GPU without DD or on a separate PME rank, and because the local state pointer
1368 // points to the global state object without DD.
1369 // FIXME: MD and EM separately set up the local state - this should happen in the same
1370 // function, which should also perform the pinning.
1371 changePinningPolicy(&globalState->x, pme_get_pinning_policy());
1374 /* Initialize the virtual site communication */
1375 vsite = initVsite(mtop, cr);
1377 calc_shifts(box, fr->shift_vec);
1379 /* With periodic molecules the charge groups should be whole at start up
1380 * and the virtual sites should not be far from their proper positions.
1382 if (!inputrec->bContinuation && MASTER(cr) && !(inputrec->ePBC != epbcNONE && inputrec->bPeriodicMols))
1384 /* Make molecules whole at start of run */
1385 if (fr->ePBC != epbcNONE)
1387 do_pbc_first_mtop(fplog, inputrec->ePBC, box, &mtop, globalState->x.rvec_array());
1391 /* Correct initial vsite positions are required
1392 * for the initial distribution in the domain decomposition
1393 * and for the initial shell prediction.
1395 constructVsitesGlobal(mtop, globalState->x);
1399 if (EEL_PME(fr->ic->eeltype) || EVDW_PME(fr->ic->vdwtype))
1401 ewaldcoeff_q = fr->ic->ewaldcoeff_q;
1402 ewaldcoeff_lj = fr->ic->ewaldcoeff_lj;
1407 /* This is a PME only node */
1409 GMX_ASSERT(globalState == nullptr,
1410 "We don't need the state on a PME only rank and expect it to be unitialized");
1412 ewaldcoeff_q = calc_ewaldcoeff_q(inputrec->rcoulomb, inputrec->ewald_rtol);
1413 ewaldcoeff_lj = calc_ewaldcoeff_lj(inputrec->rvdw, inputrec->ewald_rtol_lj);
1416 gmx_pme_t* sepPmeData = nullptr;
1417 // This reference hides the fact that PME data is owned by runner on PME-only ranks and by forcerec on other ranks
1418 GMX_ASSERT(thisRankHasDuty(cr, DUTY_PP) == (fr != nullptr),
1419 "Double-checking that only PME-only ranks have no forcerec");
1420 gmx_pme_t*& pmedata = fr ? fr->pmedata : sepPmeData;
1422 // TODO should live in ewald module once its testing is improved
1424 // Later, this program could contain kernels that might be later
1425 // re-used as auto-tuning progresses, or subsequent simulations
1427 PmeGpuProgramStorage pmeGpuProgram;
1428 if (thisRankHasPmeGpuTask)
1430 pmeGpuProgram = buildPmeGpuProgram(pmeDeviceInfo);
1433 /* Initiate PME if necessary,
1434 * either on all nodes or on dedicated PME nodes only. */
1435 if (EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype))
1437 if (mdAtoms && mdAtoms->mdatoms())
1439 nChargePerturbed = mdAtoms->mdatoms()->nChargePerturbed;
1440 if (EVDW_PME(inputrec->vdwtype))
1442 nTypePerturbed = mdAtoms->mdatoms()->nTypePerturbed;
1445 if (cr->npmenodes > 0)
1447 /* The PME only nodes need to know nChargePerturbed(FEP on Q) and nTypePerturbed(FEP on LJ)*/
1448 gmx_bcast_sim(sizeof(nChargePerturbed), &nChargePerturbed, cr);
1449 gmx_bcast_sim(sizeof(nTypePerturbed), &nTypePerturbed, cr);
1452 if (thisRankHasDuty(cr, DUTY_PME))
1456 pmedata = gmx_pme_init(cr, getNumPmeDomains(cr->dd), inputrec, nChargePerturbed != 0,
1457 nTypePerturbed != 0, mdrunOptions.reproducible, ewaldcoeff_q,
1458 ewaldcoeff_lj, gmx_omp_nthreads_get(emntPME), pmeRunMode,
1459 nullptr, pmeDeviceInfo, pmeGpuProgram.get(), mdlog);
1461 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1466 if (EI_DYNAMICS(inputrec->eI))
1468 /* Turn on signal handling on all nodes */
1470 * (A user signal from the PME nodes (if any)
1471 * is communicated to the PP nodes.
1473 signal_handler_install();
1476 pull_t* pull_work = nullptr;
1477 if (thisRankHasDuty(cr, DUTY_PP))
1479 /* Assumes uniform use of the number of OpenMP threads */
1480 walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(emntDefault));
1482 if (inputrec->bPull)
1484 /* Initialize pull code */
1485 pull_work = init_pull(fplog, inputrec->pull, inputrec, &mtop, cr, &atomSets,
1486 inputrec->fepvals->init_lambda);
1487 if (inputrec->pull->bXOutAverage || inputrec->pull->bFOutAverage)
1489 initPullHistory(pull_work, &observablesHistory);
1491 if (EI_DYNAMICS(inputrec->eI) && MASTER(cr))
1493 init_pull_output_files(pull_work, filenames.size(), filenames.data(), oenv, startingBehavior);
1497 std::unique_ptr<EnforcedRotation> enforcedRotation;
1500 /* Initialize enforced rotation code */
1502 init_rot(fplog, inputrec, filenames.size(), filenames.data(), cr, &atomSets,
1503 globalState.get(), &mtop, oenv, mdrunOptions, startingBehavior);
1506 t_swap* swap = nullptr;
1507 if (inputrec->eSwapCoords != eswapNO)
1509 /* Initialize ion swapping code */
1510 swap = init_swapcoords(fplog, inputrec,
1511 opt2fn_master("-swap", filenames.size(), filenames.data(), cr),
1512 &mtop, globalState.get(), &observablesHistory, cr, &atomSets,
1513 oenv, mdrunOptions, startingBehavior);
1516 /* Let makeConstraints know whether we have essential dynamics constraints. */
1517 auto constr = makeConstraints(mtop, *inputrec, pull_work, doEssentialDynamics, fplog,
1518 *mdAtoms->mdatoms(), cr, ms, &nrnb, wcycle, fr->bMolPBC);
1520 /* Energy terms and groups */
1521 gmx_enerdata_t enerd(mtop.groups.groups[SimulationAtomGroupType::EnergyOutput].size(),
1522 inputrec->fepvals->n_lambda);
1524 // cos acceleration is only supported by md, but older tpr
1525 // files might still combine it with other integrators
1526 GMX_RELEASE_ASSERT(inputrec->cos_accel == 0.0 || inputrec->eI == eiMD,
1527 "cos_acceleration is only supported by integrator=md");
1529 /* Kinetic energy data */
1530 gmx_ekindata_t ekind;
1531 init_ekindata(fplog, &mtop, &(inputrec->opts), &ekind, inputrec->cos_accel);
1533 /* Set up interactive MD (IMD) */
1535 makeImdSession(inputrec, cr, wcycle, &enerd, ms, &mtop, mdlog,
1536 MASTER(cr) ? globalState->x.rvec_array() : nullptr, filenames.size(),
1537 filenames.data(), oenv, mdrunOptions.imdOptions, startingBehavior);
1539 if (DOMAINDECOMP(cr))
1541 GMX_RELEASE_ASSERT(fr, "fr was NULL while cr->duty was DUTY_PP");
1542 /* This call is not included in init_domain_decomposition mainly
1543 * because fr->cginfo_mb is set later.
1545 dd_init_bondeds(fplog, cr->dd, &mtop, vsite.get(), inputrec,
1546 domdecOptions.checkBondedInteractions, fr->cginfo_mb);
1549 // TODO This is not the right place to manage the lifetime of
1550 // this data structure, but currently it's the easiest way to
1552 MdrunScheduleWorkload runScheduleWork;
1553 // Also populates the simulation constant workload description.
1554 runScheduleWork.simulationWork = createSimulationWorkload(
1555 useGpuForNonbonded, pmeRunMode, useGpuForBonded, useGpuForUpdate,
1556 devFlags.enableGpuBufferOps, devFlags.enableGpuHaloExchange,
1557 devFlags.enableGpuPmePPComm, haveEwaldSurfaceContribution(*inputrec));
1559 std::unique_ptr<gmx::StatePropagatorDataGpu> stateGpu;
1560 if (gpusWereDetected
1561 && ((useGpuForPme && thisRankHasDuty(cr, DUTY_PME))
1562 || runScheduleWork.simulationWork.useGpuBufferOps))
1564 const void* pmeStream = pme_gpu_get_device_stream(fr->pmedata);
1565 const void* localStream =
1566 fr->nbv->gpu_nbv != nullptr
1567 ? Nbnxm::gpu_get_command_stream(fr->nbv->gpu_nbv, InteractionLocality::Local)
1569 const void* nonLocalStream =
1570 fr->nbv->gpu_nbv != nullptr
1571 ? Nbnxm::gpu_get_command_stream(fr->nbv->gpu_nbv, InteractionLocality::NonLocal)
1573 const void* deviceContext = pme_gpu_get_device_context(fr->pmedata);
1574 const int paddingSize = pme_gpu_get_padding_size(fr->pmedata);
1575 GpuApiCallBehavior transferKind = (inputrec->eI == eiMD && !doRerun && !useModularSimulator)
1576 ? GpuApiCallBehavior::Async
1577 : GpuApiCallBehavior::Sync;
1579 stateGpu = std::make_unique<gmx::StatePropagatorDataGpu>(
1580 pmeStream, localStream, nonLocalStream, deviceContext, transferKind, paddingSize, wcycle);
1581 fr->stateGpu = stateGpu.get();
1584 GMX_ASSERT(stopHandlerBuilder_, "Runner must provide StopHandlerBuilder to simulator.");
1585 SimulatorBuilder simulatorBuilder;
1587 // build and run simulator object based on user-input
1588 auto simulator = simulatorBuilder.build(
1589 inputIsCompatibleWithModularSimulator, fplog, cr, ms, mdlog,
1590 static_cast<int>(filenames.size()), filenames.data(), oenv, mdrunOptions,
1591 startingBehavior, vsite.get(), constr.get(),
1592 enforcedRotation ? enforcedRotation->getLegacyEnfrot() : nullptr, deform.get(),
1593 mdModules_->outputProvider(), mdModules_->notifier(), inputrec, imdSession.get(),
1594 pull_work, swap, &mtop, fcd, globalState.get(), &observablesHistory, mdAtoms.get(),
1595 &nrnb, wcycle, fr, &enerd, &ekind, &runScheduleWork, replExParams, membed,
1596 walltime_accounting, std::move(stopHandlerBuilder_), doRerun);
1599 if (fr->pmePpCommGpu)
1601 // destroy object since it is no longer required. (This needs to be done while the GPU context still exists.)
1602 fr->pmePpCommGpu.reset();
1605 if (inputrec->bPull)
1607 finish_pull(pull_work);
1609 finish_swapcoords(swap);
1613 GMX_RELEASE_ASSERT(pmedata, "pmedata was NULL while cr->duty was not DUTY_PP");
1615 walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(emntPME));
1616 gmx_pmeonly(pmedata, cr, &nrnb, wcycle, walltime_accounting, inputrec, pmeRunMode);
1619 wallcycle_stop(wcycle, ewcRUN);
1621 /* Finish up, write some stuff
1622 * if rerunMD, don't write last frame again
1624 finish_run(fplog, mdlog, cr, inputrec, &nrnb, wcycle, walltime_accounting,
1625 fr ? fr->nbv.get() : nullptr, pmedata, EI_DYNAMICS(inputrec->eI) && !isMultiSim(ms));
1627 // clean up cycle counter
1628 wallcycle_destroy(wcycle);
1633 gmx_pme_destroy(pmedata);
1637 // FIXME: this is only here to manually unpin mdAtoms->chargeA_ and state->x,
1638 // before we destroy the GPU context(s) in free_gpu_resources().
1639 // Pinned buffers are associated with contexts in CUDA.
1640 // As soon as we destroy GPU contexts after mdrunner() exits, these lines should go.
1641 mdAtoms.reset(nullptr);
1642 globalState.reset(nullptr);
1643 mdModules_.reset(nullptr); // destruct force providers here as they might also use the GPU
1645 /* Free GPU memory and set a physical node tMPI barrier (which should eventually go away) */
1646 free_gpu_resources(fr, physicalNodeComm, hwinfo->gpu_info);
1647 free_gpu(nonbondedDeviceInfo);
1648 free_gpu(pmeDeviceInfo);
1649 done_forcerec(fr, mtop.molblock.size());
1654 free_membed(membed);
1657 /* Does what it says */
1658 print_date_and_time(fplog, cr->nodeid, "Finished mdrun", gmx_gettime());
1659 walltime_accounting_destroy(walltime_accounting);
1661 // Ensure log file content is written
1664 gmx_fio_flush(logFileHandle);
1667 /* Reset FPEs (important for unit tests) by disabling them. Assumes no
1668 * exceptions were enabled before function was called. */
1671 gmx_fedisableexcept();
1674 auto rc = static_cast<int>(gmx_get_stop_condition());
1677 /* we need to join all threads. The sub-threads join when they
1678 exit this function, but the master thread needs to be told to
1680 if (PAR(cr) && MASTER(cr))
1688 Mdrunner::~Mdrunner()
1690 // Clean up of the Manager.
1691 // This will end up getting called on every thread-MPI rank, which is unnecessary,
1692 // but okay as long as threads synchronize some time before adding or accessing
1693 // a new set of restraints.
1694 if (restraintManager_)
1696 restraintManager_->clear();
1697 GMX_ASSERT(restraintManager_->countRestraints() == 0,
1698 "restraints added during runner life time should be cleared at runner "
1703 void Mdrunner::addPotential(std::shared_ptr<gmx::IRestraintPotential> puller, const std::string& name)
1705 GMX_ASSERT(restraintManager_, "Mdrunner must have a restraint manager.");
1706 // Not sure if this should be logged through the md logger or something else,
1707 // but it is helpful to have some sort of INFO level message sent somewhere.
1708 // std::cout << "Registering restraint named " << name << std::endl;
1710 // When multiple restraints are used, it may be wasteful to register them separately.
1711 // Maybe instead register an entire Restraint Manager as a force provider.
1712 restraintManager_->addToSpec(std::move(puller), name);
1715 Mdrunner::Mdrunner(std::unique_ptr<MDModules> mdModules) : mdModules_(std::move(mdModules)) {}
1717 Mdrunner::Mdrunner(Mdrunner&&) noexcept = default;
1719 //NOLINTNEXTLINE(performance-noexcept-move-constructor) working around GCC bug 58265
1720 Mdrunner& Mdrunner::operator=(Mdrunner&& /*handle*/) noexcept(BUGFREE_NOEXCEPT_STRING) = default;
1722 class Mdrunner::BuilderImplementation
1725 BuilderImplementation() = delete;
1726 BuilderImplementation(std::unique_ptr<MDModules> mdModules, compat::not_null<SimulationContext*> context);
1727 ~BuilderImplementation();
1729 BuilderImplementation& setExtraMdrunOptions(const MdrunOptions& options,
1730 real forceWarningThreshold,
1731 StartingBehavior startingBehavior);
1733 void addDomdec(const DomdecOptions& options);
1735 void addVerletList(int nstlist);
1737 void addReplicaExchange(const ReplicaExchangeParameters& params);
1739 void addNonBonded(const char* nbpu_opt);
1741 void addPME(const char* pme_opt_, const char* pme_fft_opt_);
1743 void addBondedTaskAssignment(const char* bonded_opt);
1745 void addUpdateTaskAssignment(const char* update_opt);
1747 void addHardwareOptions(const gmx_hw_opt_t& hardwareOptions);
1749 void addFilenames(ArrayRef<const t_filenm> filenames);
1751 void addOutputEnvironment(gmx_output_env_t* outputEnvironment);
1753 void addLogFile(t_fileio* logFileHandle);
1755 void addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder);
1760 // Default parameters copied from runner.h
1761 // \todo Clarify source(s) of default parameters.
1763 const char* nbpu_opt_ = nullptr;
1764 const char* pme_opt_ = nullptr;
1765 const char* pme_fft_opt_ = nullptr;
1766 const char* bonded_opt_ = nullptr;
1767 const char* update_opt_ = nullptr;
1769 MdrunOptions mdrunOptions_;
1771 DomdecOptions domdecOptions_;
1773 ReplicaExchangeParameters replicaExchangeParameters_;
1775 //! Command-line override for the duration of a neighbor list with the Verlet scheme.
1778 //! Multisim communicator handle.
1779 gmx_multisim_t* multiSimulation_;
1781 //! mdrun communicator
1782 MPI_Comm communicator_ = MPI_COMM_NULL;
1784 //! Print a warning if any force is larger than this (in kJ/mol nm).
1785 real forceWarningThreshold_ = -1;
1787 //! Whether the simulation will start afresh, or restart with/without appending.
1788 StartingBehavior startingBehavior_ = StartingBehavior::NewSimulation;
1790 //! The modules that comprise the functionality of mdrun.
1791 std::unique_ptr<MDModules> mdModules_;
1793 //! \brief Parallelism information.
1794 gmx_hw_opt_t hardwareOptions_;
1796 //! filename options for simulation.
1797 ArrayRef<const t_filenm> filenames_;
1799 /*! \brief Handle to output environment.
1801 * \todo gmx_output_env_t needs lifetime management.
1803 gmx_output_env_t* outputEnvironment_ = nullptr;
1805 /*! \brief Non-owning handle to MD log file.
1807 * \todo Context should own output facilities for client.
1808 * \todo Improve log file handle management.
1810 * Code managing the FILE* relies on the ability to set it to
1811 * nullptr to check whether the filehandle is valid.
1813 t_fileio* logFileHandle_ = nullptr;
1816 * \brief Builder for simulation stop signal handler.
1818 std::unique_ptr<StopHandlerBuilder> stopHandlerBuilder_ = nullptr;
1821 Mdrunner::BuilderImplementation::BuilderImplementation(std::unique_ptr<MDModules> mdModules,
1822 compat::not_null<SimulationContext*> context) :
1823 mdModules_(std::move(mdModules))
1825 communicator_ = context->communicator_;
1826 multiSimulation_ = context->multiSimulation_.get();
1829 Mdrunner::BuilderImplementation::~BuilderImplementation() = default;
1831 Mdrunner::BuilderImplementation&
1832 Mdrunner::BuilderImplementation::setExtraMdrunOptions(const MdrunOptions& options,
1833 const real forceWarningThreshold,
1834 const StartingBehavior startingBehavior)
1836 mdrunOptions_ = options;
1837 forceWarningThreshold_ = forceWarningThreshold;
1838 startingBehavior_ = startingBehavior;
1842 void Mdrunner::BuilderImplementation::addDomdec(const DomdecOptions& options)
1844 domdecOptions_ = options;
1847 void Mdrunner::BuilderImplementation::addVerletList(int nstlist)
1852 void Mdrunner::BuilderImplementation::addReplicaExchange(const ReplicaExchangeParameters& params)
1854 replicaExchangeParameters_ = params;
1857 Mdrunner Mdrunner::BuilderImplementation::build()
1859 auto newRunner = Mdrunner(std::move(mdModules_));
1861 newRunner.mdrunOptions = mdrunOptions_;
1862 newRunner.pforce = forceWarningThreshold_;
1863 newRunner.startingBehavior = startingBehavior_;
1864 newRunner.domdecOptions = domdecOptions_;
1866 // \todo determine an invariant to check or confirm that all gmx_hw_opt_t objects are valid
1867 newRunner.hw_opt = hardwareOptions_;
1869 // No invariant to check. This parameter exists to optionally override other behavior.
1870 newRunner.nstlist_cmdline = nstlist_;
1872 newRunner.replExParams = replicaExchangeParameters_;
1874 newRunner.filenames = filenames_;
1876 newRunner.communicator = communicator_;
1878 // nullptr is a valid value for the multisim handle
1879 newRunner.ms = multiSimulation_;
1881 // \todo Clarify ownership and lifetime management for gmx_output_env_t
1882 // \todo Update sanity checking when output environment has clearly specified invariants.
1883 // Initialization and default values for oenv are not well specified in the current version.
1884 if (outputEnvironment_)
1886 newRunner.oenv = outputEnvironment_;
1890 GMX_THROW(gmx::APIError(
1891 "MdrunnerBuilder::addOutputEnvironment() is required before build()"));
1894 newRunner.logFileHandle = logFileHandle_;
1898 newRunner.nbpu_opt = nbpu_opt_;
1902 GMX_THROW(gmx::APIError("MdrunnerBuilder::addNonBonded() is required before build()"));
1905 if (pme_opt_ && pme_fft_opt_)
1907 newRunner.pme_opt = pme_opt_;
1908 newRunner.pme_fft_opt = pme_fft_opt_;
1912 GMX_THROW(gmx::APIError("MdrunnerBuilder::addElectrostatics() is required before build()"));
1917 newRunner.bonded_opt = bonded_opt_;
1921 GMX_THROW(gmx::APIError(
1922 "MdrunnerBuilder::addBondedTaskAssignment() is required before build()"));
1927 newRunner.update_opt = update_opt_;
1931 GMX_THROW(gmx::APIError(
1932 "MdrunnerBuilder::addUpdateTaskAssignment() is required before build() "));
1936 newRunner.restraintManager_ = std::make_unique<gmx::RestraintManager>();
1938 if (stopHandlerBuilder_)
1940 newRunner.stopHandlerBuilder_ = std::move(stopHandlerBuilder_);
1944 newRunner.stopHandlerBuilder_ = std::make_unique<StopHandlerBuilder>();
1950 void Mdrunner::BuilderImplementation::addNonBonded(const char* nbpu_opt)
1952 nbpu_opt_ = nbpu_opt;
1955 void Mdrunner::BuilderImplementation::addPME(const char* pme_opt, const char* pme_fft_opt)
1958 pme_fft_opt_ = pme_fft_opt;
1961 void Mdrunner::BuilderImplementation::addBondedTaskAssignment(const char* bonded_opt)
1963 bonded_opt_ = bonded_opt;
1966 void Mdrunner::BuilderImplementation::addUpdateTaskAssignment(const char* update_opt)
1968 update_opt_ = update_opt;
1971 void Mdrunner::BuilderImplementation::addHardwareOptions(const gmx_hw_opt_t& hardwareOptions)
1973 hardwareOptions_ = hardwareOptions;
1976 void Mdrunner::BuilderImplementation::addFilenames(ArrayRef<const t_filenm> filenames)
1978 filenames_ = filenames;
1981 void Mdrunner::BuilderImplementation::addOutputEnvironment(gmx_output_env_t* outputEnvironment)
1983 outputEnvironment_ = outputEnvironment;
1986 void Mdrunner::BuilderImplementation::addLogFile(t_fileio* logFileHandle)
1988 logFileHandle_ = logFileHandle;
1991 void Mdrunner::BuilderImplementation::addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder)
1993 stopHandlerBuilder_ = std::move(builder);
1996 MdrunnerBuilder::MdrunnerBuilder(std::unique_ptr<MDModules> mdModules,
1997 compat::not_null<SimulationContext*> context) :
1998 impl_{ std::make_unique<Mdrunner::BuilderImplementation>(std::move(mdModules), context) }
2002 MdrunnerBuilder::~MdrunnerBuilder() = default;
2004 MdrunnerBuilder& MdrunnerBuilder::addSimulationMethod(const MdrunOptions& options,
2005 real forceWarningThreshold,
2006 const StartingBehavior startingBehavior)
2008 impl_->setExtraMdrunOptions(options, forceWarningThreshold, startingBehavior);
2012 MdrunnerBuilder& MdrunnerBuilder::addDomainDecomposition(const DomdecOptions& options)
2014 impl_->addDomdec(options);
2018 MdrunnerBuilder& MdrunnerBuilder::addNeighborList(int nstlist)
2020 impl_->addVerletList(nstlist);
2024 MdrunnerBuilder& MdrunnerBuilder::addReplicaExchange(const ReplicaExchangeParameters& params)
2026 impl_->addReplicaExchange(params);
2030 MdrunnerBuilder& MdrunnerBuilder::addNonBonded(const char* nbpu_opt)
2032 impl_->addNonBonded(nbpu_opt);
2036 MdrunnerBuilder& MdrunnerBuilder::addElectrostatics(const char* pme_opt, const char* pme_fft_opt)
2038 // The builder method may become more general in the future, but in this version,
2039 // parameters for PME electrostatics are both required and the only parameters
2041 if (pme_opt && pme_fft_opt)
2043 impl_->addPME(pme_opt, pme_fft_opt);
2048 gmx::InvalidInputError("addElectrostatics() arguments must be non-null pointers."));
2053 MdrunnerBuilder& MdrunnerBuilder::addBondedTaskAssignment(const char* bonded_opt)
2055 impl_->addBondedTaskAssignment(bonded_opt);
2059 MdrunnerBuilder& MdrunnerBuilder::addUpdateTaskAssignment(const char* update_opt)
2061 impl_->addUpdateTaskAssignment(update_opt);
2065 Mdrunner MdrunnerBuilder::build()
2067 return impl_->build();
2070 MdrunnerBuilder& MdrunnerBuilder::addHardwareOptions(const gmx_hw_opt_t& hardwareOptions)
2072 impl_->addHardwareOptions(hardwareOptions);
2076 MdrunnerBuilder& MdrunnerBuilder::addFilenames(ArrayRef<const t_filenm> filenames)
2078 impl_->addFilenames(filenames);
2082 MdrunnerBuilder& MdrunnerBuilder::addOutputEnvironment(gmx_output_env_t* outputEnvironment)
2084 impl_->addOutputEnvironment(outputEnvironment);
2088 MdrunnerBuilder& MdrunnerBuilder::addLogFile(t_fileio* logFileHandle)
2090 impl_->addLogFile(logFileHandle);
2094 MdrunnerBuilder& MdrunnerBuilder::addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder)
2096 impl_->addStopHandlerBuilder(std::move(builder));
2100 MdrunnerBuilder::MdrunnerBuilder(MdrunnerBuilder&&) noexcept = default;
2102 MdrunnerBuilder& MdrunnerBuilder::operator=(MdrunnerBuilder&&) noexcept = default;