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, 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/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/simulation_workload.h"
118 #include "gromacs/mdtypes/state.h"
119 #include "gromacs/mdtypes/state_propagator_data_gpu.h"
120 #include "gromacs/nbnxm/gpu_data_mgmt.h"
121 #include "gromacs/nbnxm/nbnxm.h"
122 #include "gromacs/nbnxm/pairlist_tuning.h"
123 #include "gromacs/pbcutil/pbc.h"
124 #include "gromacs/pulling/output.h"
125 #include "gromacs/pulling/pull.h"
126 #include "gromacs/pulling/pull_rotation.h"
127 #include "gromacs/restraint/manager.h"
128 #include "gromacs/restraint/restraintmdmodule.h"
129 #include "gromacs/restraint/restraintpotential.h"
130 #include "gromacs/swap/swapcoords.h"
131 #include "gromacs/taskassignment/decidegpuusage.h"
132 #include "gromacs/taskassignment/decidesimulationworkload.h"
133 #include "gromacs/taskassignment/resourcedivision.h"
134 #include "gromacs/taskassignment/taskassignment.h"
135 #include "gromacs/taskassignment/usergpuids.h"
136 #include "gromacs/timing/gpu_timing.h"
137 #include "gromacs/timing/wallcycle.h"
138 #include "gromacs/timing/wallcyclereporting.h"
139 #include "gromacs/topology/mtop_util.h"
140 #include "gromacs/trajectory/trajectoryframe.h"
141 #include "gromacs/utility/basenetwork.h"
142 #include "gromacs/utility/cstringutil.h"
143 #include "gromacs/utility/exceptions.h"
144 #include "gromacs/utility/fatalerror.h"
145 #include "gromacs/utility/filestream.h"
146 #include "gromacs/utility/gmxassert.h"
147 #include "gromacs/utility/gmxmpi.h"
148 #include "gromacs/utility/keyvaluetree.h"
149 #include "gromacs/utility/logger.h"
150 #include "gromacs/utility/loggerbuilder.h"
151 #include "gromacs/utility/mdmodulenotification.h"
152 #include "gromacs/utility/physicalnodecommunicator.h"
153 #include "gromacs/utility/pleasecite.h"
154 #include "gromacs/utility/programcontext.h"
155 #include "gromacs/utility/smalloc.h"
156 #include "gromacs/utility/stringutil.h"
158 #include "isimulator.h"
159 #include "replicaexchange.h"
160 #include "simulatorbuilder.h"
163 # include "corewrap.h"
169 /*! \brief Structure that holds boolean flags corresponding to the development
170 * features present enabled through environment variables.
173 struct DevelopmentFeatureFlags
175 //! True if the Buffer ops development feature is enabled
176 // TODO: when the trigger of the buffer ops offload is fully automated this should go away
177 bool enableGpuBufferOps = false;
178 //! True if the GPU halo exchange development feature is enabled
179 bool enableGpuHaloExchange = false;
180 //! True if the PME PP direct communication GPU development feature is enabled
181 bool enableGpuPmePPComm = false;
184 /*! \brief Manage any development feature flag variables encountered
186 * The use of dev features indicated by environment variables is
187 * logged in order to ensure that runs with such features enabled can
188 * be identified from their log and standard output. Any cross
189 * dependencies are also checked, and if unsatisfied, a fatal error
192 * Note that some development features overrides are applied already here:
193 * the GPU communication flags are set to false in non-tMPI and non-CUDA builds.
195 * \param[in] mdlog Logger object.
196 * \param[in] useGpuForNonbonded True if the nonbonded task is offloaded in this run.
197 * \param[in] pmeRunMode The PME run mode for this run
198 * \returns The object populated with development feature flags.
200 static DevelopmentFeatureFlags manageDevelopmentFeatures(const gmx::MDLogger& mdlog,
201 const bool useGpuForNonbonded,
202 const PmeRunMode pmeRunMode)
204 DevelopmentFeatureFlags devFlags;
206 // Some builds of GCC 5 give false positive warnings that these
207 // getenv results are ignored when clearly they are used.
208 #pragma GCC diagnostic push
209 #pragma GCC diagnostic ignored "-Wunused-result"
210 devFlags.enableGpuBufferOps = (getenv("GMX_USE_GPU_BUFFER_OPS") != nullptr)
211 && (GMX_GPU == GMX_GPU_CUDA) && useGpuForNonbonded;
212 devFlags.enableGpuHaloExchange =
213 (getenv("GMX_GPU_DD_COMMS") != nullptr && GMX_THREAD_MPI && (GMX_GPU == GMX_GPU_CUDA));
214 devFlags.enableGpuPmePPComm =
215 (getenv("GMX_GPU_PME_PP_COMMS") != nullptr && GMX_THREAD_MPI && (GMX_GPU == GMX_GPU_CUDA));
216 #pragma GCC diagnostic pop
218 if (devFlags.enableGpuBufferOps)
220 GMX_LOG(mdlog.warning)
222 .appendTextFormatted(
223 "NOTE: This run uses the 'GPU buffer ops' feature, enabled by the "
224 "GMX_USE_GPU_BUFFER_OPS environment variable.");
227 if (devFlags.enableGpuHaloExchange)
229 if (useGpuForNonbonded)
231 if (!devFlags.enableGpuBufferOps)
234 "Cannot enable GPU halo exchange without GPU buffer operations, set "
235 "GMX_USE_GPU_BUFFER_OPS=1\n");
237 GMX_LOG(mdlog.warning)
239 .appendTextFormatted(
240 "NOTE: This run uses the 'GPU halo exchange' feature, enabled by the "
241 "GMX_GPU_DD_COMMS environment variable.");
245 GMX_LOG(mdlog.warning)
247 .appendTextFormatted(
248 "NOTE: GMX_GPU_DD_COMMS environment variable detected, but the 'GPU "
249 "halo exchange' feature will not be enabled as nonbonded interactions "
250 "are not offloaded.");
251 devFlags.enableGpuHaloExchange = false;
255 if (devFlags.enableGpuPmePPComm)
257 if (pmeRunMode == PmeRunMode::GPU)
259 GMX_LOG(mdlog.warning)
261 .appendTextFormatted(
262 "NOTE: This run uses the 'GPU PME-PP communications' feature, enabled "
263 "by the GMX_GPU_PME_PP_COMMS environment variable.");
267 std::string clarification;
268 if (pmeRunMode == PmeRunMode::Mixed)
271 "PME FFT and gather are not offloaded to the GPU (PME is running in mixed "
276 clarification = "PME is not offloaded to the GPU.";
278 GMX_LOG(mdlog.warning)
281 "NOTE: GMX_GPU_PME_PP_COMMS environment variable detected, but the "
282 "'GPU PME-PP communications' feature was not enabled as "
284 devFlags.enableGpuPmePPComm = false;
291 /*! \brief Barrier for safe simultaneous thread access to mdrunner data
293 * Used to ensure that the master thread does not modify mdrunner during copy
294 * on the spawned threads. */
295 static void threadMpiMdrunnerAccessBarrier()
298 MPI_Barrier(MPI_COMM_WORLD);
302 Mdrunner Mdrunner::cloneOnSpawnedThread() const
304 auto newRunner = Mdrunner(std::make_unique<MDModules>());
306 // All runners in the same process share a restraint manager resource because it is
307 // part of the interface to the client code, which is associated only with the
308 // original thread. Handles to the same resources can be obtained by copy.
310 newRunner.restraintManager_ = std::make_unique<RestraintManager>(*restraintManager_);
313 // Copy members of master runner.
314 // \todo Replace with builder when Simulation context and/or runner phases are better defined.
315 // Ref https://redmine.gromacs.org/issues/2587 and https://redmine.gromacs.org/issues/2375
316 newRunner.hw_opt = hw_opt;
317 newRunner.filenames = filenames;
319 newRunner.oenv = oenv;
320 newRunner.mdrunOptions = mdrunOptions;
321 newRunner.domdecOptions = domdecOptions;
322 newRunner.nbpu_opt = nbpu_opt;
323 newRunner.pme_opt = pme_opt;
324 newRunner.pme_fft_opt = pme_fft_opt;
325 newRunner.bonded_opt = bonded_opt;
326 newRunner.update_opt = update_opt;
327 newRunner.nstlist_cmdline = nstlist_cmdline;
328 newRunner.replExParams = replExParams;
329 newRunner.pforce = pforce;
330 // Give the spawned thread the newly created valid communicator
331 // for the simulation.
332 newRunner.communicator = MPI_COMM_WORLD;
334 newRunner.startingBehavior = startingBehavior;
335 newRunner.stopHandlerBuilder_ = std::make_unique<StopHandlerBuilder>(*stopHandlerBuilder_);
337 threadMpiMdrunnerAccessBarrier();
342 /*! \brief The callback used for running on spawned threads.
344 * Obtains the pointer to the master mdrunner object from the one
345 * argument permitted to the thread-launch API call, copies it to make
346 * a new runner for this thread, reinitializes necessary data, and
347 * proceeds to the simulation. */
348 static void mdrunner_start_fn(const void* arg)
352 auto masterMdrunner = reinterpret_cast<const gmx::Mdrunner*>(arg);
353 /* copy the arg list to make sure that it's thread-local. This
354 doesn't copy pointed-to items, of course; fnm, cr and fplog
355 are reset in the call below, all others should be const. */
356 gmx::Mdrunner mdrunner = masterMdrunner->cloneOnSpawnedThread();
359 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
363 void Mdrunner::spawnThreads(int numThreadsToLaunch)
366 /* now spawn new threads that start mdrunner_start_fn(), while
367 the main thread returns. Thread affinity is handled later. */
368 if (tMPI_Init_fn(TRUE, numThreadsToLaunch, TMPI_AFFINITY_NONE, mdrunner_start_fn,
369 static_cast<const void*>(this))
372 GMX_THROW(gmx::InternalError("Failed to spawn thread-MPI threads"));
375 // Give the master thread the newly created valid communicator for
377 communicator = MPI_COMM_WORLD;
378 threadMpiMdrunnerAccessBarrier();
380 GMX_UNUSED_VALUE(numThreadsToLaunch);
381 GMX_UNUSED_VALUE(mdrunner_start_fn);
387 /*! \brief Initialize variables for Verlet scheme simulation */
388 static void prepare_verlet_scheme(FILE* fplog,
392 const gmx_mtop_t* mtop,
394 bool makeGpuPairList,
395 const gmx::CpuInfo& cpuinfo)
397 /* For NVE simulations, we will retain the initial list buffer */
398 if (EI_DYNAMICS(ir->eI) && ir->verletbuf_tol > 0 && !(EI_MD(ir->eI) && ir->etc == etcNO))
400 /* Update the Verlet buffer size for the current run setup */
402 /* Here we assume SIMD-enabled kernels are being used. But as currently
403 * calc_verlet_buffer_size gives the same results for 4x8 and 4x4
404 * and 4x2 gives a larger buffer than 4x4, this is ok.
406 ListSetupType listType =
407 (makeGpuPairList ? ListSetupType::Gpu : ListSetupType::CpuSimdWhenSupported);
408 VerletbufListSetup listSetup = verletbufGetSafeListSetup(listType);
410 const real rlist_new =
411 calcVerletBufferSize(*mtop, det(box), *ir, ir->nstlist, ir->nstlist - 1, -1, listSetup);
413 if (rlist_new != ir->rlist)
415 if (fplog != nullptr)
418 "\nChanging rlist from %g to %g for non-bonded %dx%d atom kernels\n\n",
419 ir->rlist, rlist_new, listSetup.cluster_size_i, listSetup.cluster_size_j);
421 ir->rlist = rlist_new;
425 if (nstlist_cmdline > 0 && (!EI_DYNAMICS(ir->eI) || ir->verletbuf_tol <= 0))
427 gmx_fatal(FARGS, "Can not set nstlist without %s",
428 !EI_DYNAMICS(ir->eI) ? "dynamics" : "verlet-buffer-tolerance");
431 if (EI_DYNAMICS(ir->eI))
433 /* Set or try nstlist values */
434 increaseNstlist(fplog, cr, ir, nstlist_cmdline, mtop, box, makeGpuPairList, cpuinfo);
438 /*! \brief Override the nslist value in inputrec
440 * with value passed on the command line (if any)
442 static void override_nsteps_cmdline(const gmx::MDLogger& mdlog, int64_t nsteps_cmdline, t_inputrec* ir)
446 /* override with anything else than the default -2 */
447 if (nsteps_cmdline > -2)
449 char sbuf_steps[STEPSTRSIZE];
450 char sbuf_msg[STRLEN];
452 ir->nsteps = nsteps_cmdline;
453 if (EI_DYNAMICS(ir->eI) && nsteps_cmdline != -1)
456 "Overriding nsteps with value passed on the command line: %s steps, %.3g ps",
457 gmx_step_str(nsteps_cmdline, sbuf_steps), fabs(nsteps_cmdline * ir->delta_t));
461 sprintf(sbuf_msg, "Overriding nsteps with value passed on the command line: %s steps",
462 gmx_step_str(nsteps_cmdline, sbuf_steps));
465 GMX_LOG(mdlog.warning).asParagraph().appendText(sbuf_msg);
467 else if (nsteps_cmdline < -2)
469 gmx_fatal(FARGS, "Invalid nsteps value passed on the command line: %" PRId64, nsteps_cmdline);
471 /* Do nothing if nsteps_cmdline == -2 */
477 /*! \brief Return whether GPU acceleration of nonbondeds is supported with the given settings.
479 * If not, and if a warning may be issued, logs a warning about
480 * falling back to CPU code. With thread-MPI, only the first
481 * call to this function should have \c issueWarning true. */
482 static bool gpuAccelerationOfNonbondedIsUseful(const MDLogger& mdlog, const t_inputrec& ir, bool issueWarning)
484 bool gpuIsUseful = true;
487 if (ir.opts.ngener - ir.nwall > 1)
489 /* The GPU code does not support more than one energy group.
490 * If the user requested GPUs explicitly, a fatal error is given later.
494 "Multiple energy groups is not implemented for GPUs, falling back to the CPU. "
495 "For better performance, run on the GPU without energy groups and then do "
496 "gmx mdrun -rerun option on the trajectory with an energy group .tpr file.";
502 warning = "TPI is not implemented for GPUs.";
505 if (!gpuIsUseful && issueWarning)
507 GMX_LOG(mdlog.warning).asParagraph().appendText(warning);
513 //! Initializes the logger for mdrun.
514 static gmx::LoggerOwner buildLogger(FILE* fplog, const bool isSimulationMasterRank)
516 gmx::LoggerBuilder builder;
517 if (fplog != nullptr)
519 builder.addTargetFile(gmx::MDLogger::LogLevel::Info, fplog);
521 if (isSimulationMasterRank)
523 builder.addTargetStream(gmx::MDLogger::LogLevel::Warning, &gmx::TextOutputFile::standardError());
525 return builder.build();
528 //! Make a TaskTarget from an mdrun argument string.
529 static TaskTarget findTaskTarget(const char* optionString)
531 TaskTarget returnValue = TaskTarget::Auto;
533 if (strncmp(optionString, "auto", 3) == 0)
535 returnValue = TaskTarget::Auto;
537 else if (strncmp(optionString, "cpu", 3) == 0)
539 returnValue = TaskTarget::Cpu;
541 else if (strncmp(optionString, "gpu", 3) == 0)
543 returnValue = TaskTarget::Gpu;
547 GMX_ASSERT(false, "Option string should have been checked for sanity already");
553 //! Finish run, aggregate data to print performance info.
554 static void finish_run(FILE* fplog,
555 const gmx::MDLogger& mdlog,
557 const t_inputrec* inputrec,
559 gmx_wallcycle_t wcycle,
560 gmx_walltime_accounting_t walltime_accounting,
561 nonbonded_verlet_t* nbv,
562 const gmx_pme_t* pme,
566 double nbfs = 0, mflop = 0;
567 double elapsed_time, elapsed_time_over_all_ranks, elapsed_time_over_all_threads,
568 elapsed_time_over_all_threads_over_all_ranks;
569 /* Control whether it is valid to print a report. Only the
570 simulation master may print, but it should not do so if the run
571 terminated e.g. before a scheduled reset step. This is
572 complicated by the fact that PME ranks are unaware of the
573 reason why they were sent a pmerecvqxFINISH. To avoid
574 communication deadlocks, we always do the communication for the
575 report, even if we've decided not to write the report, because
576 how long it takes to finish the run is not important when we've
577 decided not to report on the simulation performance.
579 Further, we only report performance for dynamical integrators,
580 because those are the only ones for which we plan to
581 consider doing any optimizations. */
582 bool printReport = EI_DYNAMICS(inputrec->eI) && SIMMASTER(cr);
584 if (printReport && !walltime_accounting_get_valid_finish(walltime_accounting))
586 GMX_LOG(mdlog.warning)
588 .appendText("Simulation ended prematurely, no performance report will be written.");
593 std::unique_ptr<t_nrnb> nrnbTotalStorage;
596 nrnbTotalStorage = std::make_unique<t_nrnb>();
597 nrnb_tot = nrnbTotalStorage.get();
599 MPI_Allreduce(nrnb->n, nrnb_tot->n, eNRNB, MPI_DOUBLE, MPI_SUM, cr->mpi_comm_mysim);
607 elapsed_time = walltime_accounting_get_time_since_reset(walltime_accounting);
608 elapsed_time_over_all_threads =
609 walltime_accounting_get_time_since_reset_over_all_threads(walltime_accounting);
613 /* reduce elapsed_time over all MPI ranks in the current simulation */
614 MPI_Allreduce(&elapsed_time, &elapsed_time_over_all_ranks, 1, MPI_DOUBLE, MPI_SUM,
616 elapsed_time_over_all_ranks /= cr->nnodes;
617 /* Reduce elapsed_time_over_all_threads over all MPI ranks in the
618 * current simulation. */
619 MPI_Allreduce(&elapsed_time_over_all_threads, &elapsed_time_over_all_threads_over_all_ranks,
620 1, MPI_DOUBLE, MPI_SUM, cr->mpi_comm_mysim);
625 elapsed_time_over_all_ranks = elapsed_time;
626 elapsed_time_over_all_threads_over_all_ranks = elapsed_time_over_all_threads;
631 print_flop(fplog, nrnb_tot, &nbfs, &mflop);
634 if (thisRankHasDuty(cr, DUTY_PP) && DOMAINDECOMP(cr))
636 print_dd_statistics(cr, inputrec, fplog);
639 /* TODO Move the responsibility for any scaling by thread counts
640 * to the code that handled the thread region, so that there's a
641 * mechanism to keep cycle counting working during the transition
642 * to task parallelism. */
643 int nthreads_pp = gmx_omp_nthreads_get(emntNonbonded);
644 int nthreads_pme = gmx_omp_nthreads_get(emntPME);
645 wallcycle_scale_by_num_threads(wcycle, thisRankHasDuty(cr, DUTY_PME) && !thisRankHasDuty(cr, DUTY_PP),
646 nthreads_pp, nthreads_pme);
647 auto cycle_sum(wallcycle_sum(cr, wcycle));
651 auto nbnxn_gpu_timings =
652 (nbv != nullptr && nbv->useGpu()) ? Nbnxm::gpu_get_timings(nbv->gpu_nbv) : nullptr;
653 gmx_wallclock_gpu_pme_t pme_gpu_timings = {};
655 if (pme_gpu_task_enabled(pme))
657 pme_gpu_get_timings(pme, &pme_gpu_timings);
659 wallcycle_print(fplog, mdlog, cr->nnodes, cr->npmenodes, nthreads_pp, nthreads_pme,
660 elapsed_time_over_all_ranks, wcycle, cycle_sum, nbnxn_gpu_timings,
663 if (EI_DYNAMICS(inputrec->eI))
665 delta_t = inputrec->delta_t;
670 print_perf(fplog, elapsed_time_over_all_threads_over_all_ranks, elapsed_time_over_all_ranks,
671 walltime_accounting_get_nsteps_done_since_reset(walltime_accounting),
672 delta_t, nbfs, mflop);
676 print_perf(stderr, elapsed_time_over_all_threads_over_all_ranks, elapsed_time_over_all_ranks,
677 walltime_accounting_get_nsteps_done_since_reset(walltime_accounting),
678 delta_t, nbfs, mflop);
683 int Mdrunner::mdrunner()
686 t_forcerec* fr = nullptr;
687 t_fcdata* fcd = nullptr;
688 real ewaldcoeff_q = 0;
689 real ewaldcoeff_lj = 0;
690 int nChargePerturbed = -1, nTypePerturbed = 0;
691 gmx_wallcycle_t wcycle;
692 gmx_walltime_accounting_t walltime_accounting = nullptr;
693 gmx_membed_t* membed = nullptr;
694 gmx_hw_info_t* hwinfo = nullptr;
696 /* CAUTION: threads may be started later on in this function, so
697 cr doesn't reflect the final parallel state right now */
700 /* TODO: inputrec should tell us whether we use an algorithm, not a file option */
701 const bool doEssentialDynamics = opt2bSet("-ei", filenames.size(), filenames.data());
702 const bool doMembed = opt2bSet("-membed", filenames.size(), filenames.data());
703 const bool doRerun = mdrunOptions.rerun;
705 // Handle task-assignment related user options.
706 EmulateGpuNonbonded emulateGpuNonbonded =
707 (getenv("GMX_EMULATE_GPU") != nullptr ? EmulateGpuNonbonded::Yes : EmulateGpuNonbonded::No);
709 std::vector<int> userGpuTaskAssignment;
712 userGpuTaskAssignment = parseUserTaskAssignmentString(hw_opt.userGpuTaskAssignment);
714 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
715 auto nonbondedTarget = findTaskTarget(nbpu_opt);
716 auto pmeTarget = findTaskTarget(pme_opt);
717 auto pmeFftTarget = findTaskTarget(pme_fft_opt);
718 auto bondedTarget = findTaskTarget(bonded_opt);
719 auto updateTarget = findTaskTarget(update_opt);
720 PmeRunMode pmeRunMode = PmeRunMode::None;
722 FILE* fplog = nullptr;
723 // If we are appending, we don't write log output because we need
724 // to check that the old log file matches what the checkpoint file
725 // expects. Otherwise, we should start to write log output now if
726 // there is a file ready for it.
727 if (logFileHandle != nullptr && startingBehavior != StartingBehavior::RestartWithAppending)
729 fplog = gmx_fio_getfp(logFileHandle);
731 const bool isSimulationMasterRank = findIsSimulationMasterRank(ms, communicator);
732 gmx::LoggerOwner logOwner(buildLogger(fplog, isSimulationMasterRank));
733 gmx::MDLogger mdlog(logOwner.logger());
735 // TODO The thread-MPI master rank makes a working
736 // PhysicalNodeCommunicator here, but it gets rebuilt by all ranks
737 // after the threads have been launched. This works because no use
738 // is made of that communicator until after the execution paths
739 // have rejoined. But it is likely that we can improve the way
740 // this is expressed, e.g. by expressly running detection only the
741 // master rank for thread-MPI, rather than relying on the mutex
742 // and reference count.
743 PhysicalNodeCommunicator physicalNodeComm(communicator, gmx_physicalnode_id_hash());
744 hwinfo = gmx_detect_hardware(mdlog, physicalNodeComm);
746 gmx_print_detected_hardware(fplog, isSimulationMasterRank && isMasterSim(ms), mdlog, hwinfo);
748 std::vector<int> gpuIdsToUse = makeGpuIdsToUse(hwinfo->gpu_info, hw_opt.gpuIdsAvailable);
750 // Print citation requests after all software/hardware printing
751 pleaseCiteGromacs(fplog);
753 // TODO Replace this by unique_ptr once t_inputrec is C++
754 t_inputrec inputrecInstance;
755 t_inputrec* inputrec = nullptr;
756 std::unique_ptr<t_state> globalState;
758 auto partialDeserializedTpr = std::make_unique<PartialDeserializedTprFile>();
760 if (isSimulationMasterRank)
762 /* Only the master rank has the global state */
763 globalState = std::make_unique<t_state>();
765 /* Read (nearly) all data required for the simulation
766 * and keep the partly serialized tpr contents to send to other ranks later
768 *partialDeserializedTpr = read_tpx_state(ftp2fn(efTPR, filenames.size(), filenames.data()),
769 &inputrecInstance, globalState.get(), &mtop);
770 inputrec = &inputrecInstance;
773 /* Check and update the hardware options for internal consistency */
774 checkAndUpdateHardwareOptions(mdlog, &hw_opt, isSimulationMasterRank, domdecOptions.numPmeRanks,
777 if (GMX_THREAD_MPI && isSimulationMasterRank)
779 bool useGpuForNonbonded = false;
780 bool useGpuForPme = false;
783 GMX_RELEASE_ASSERT(inputrec != nullptr, "Keep the compiler happy");
785 // If the user specified the number of ranks, then we must
786 // respect that, but in default mode, we need to allow for
787 // the number of GPUs to choose the number of ranks.
788 auto canUseGpuForNonbonded = buildSupportsNonbondedOnGpu(nullptr);
789 useGpuForNonbonded = decideWhetherToUseGpusForNonbondedWithThreadMpi(
790 nonbondedTarget, gpuIdsToUse, userGpuTaskAssignment, emulateGpuNonbonded,
791 canUseGpuForNonbonded,
792 gpuAccelerationOfNonbondedIsUseful(mdlog, *inputrec, GMX_THREAD_MPI),
793 hw_opt.nthreads_tmpi);
794 useGpuForPme = decideWhetherToUseGpusForPmeWithThreadMpi(
795 useGpuForNonbonded, pmeTarget, gpuIdsToUse, userGpuTaskAssignment, *hwinfo,
796 *inputrec, mtop, hw_opt.nthreads_tmpi, domdecOptions.numPmeRanks);
798 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
800 /* Determine how many thread-MPI ranks to start.
802 * TODO Over-writing the user-supplied value here does
803 * prevent any possible subsequent checks from working
805 hw_opt.nthreads_tmpi = get_nthreads_mpi(hwinfo, &hw_opt, gpuIdsToUse, useGpuForNonbonded,
806 useGpuForPme, inputrec, &mtop, mdlog, doMembed);
808 // Now start the threads for thread MPI.
809 spawnThreads(hw_opt.nthreads_tmpi);
810 // The spawned threads enter mdrunner() and execution of
811 // master and spawned threads joins at the end of this block.
812 physicalNodeComm = PhysicalNodeCommunicator(communicator, gmx_physicalnode_id_hash());
815 GMX_RELEASE_ASSERT(communicator == MPI_COMM_WORLD, "Must have valid world communicator");
816 CommrecHandle crHandle = init_commrec(communicator, ms);
817 t_commrec* cr = crHandle.get();
818 GMX_RELEASE_ASSERT(cr != nullptr, "Must have valid commrec");
822 /* now broadcast everything to the non-master nodes/threads: */
823 if (!isSimulationMasterRank)
825 inputrec = &inputrecInstance;
827 init_parallel(cr, inputrec, &mtop, partialDeserializedTpr.get());
829 GMX_RELEASE_ASSERT(inputrec != nullptr, "All ranks should have a valid inputrec now");
830 partialDeserializedTpr.reset(nullptr);
832 // Now the number of ranks is known to all ranks, and each knows
833 // the inputrec read by the master rank. The ranks can now all run
834 // the task-deciding functions and will agree on the result
835 // without needing to communicate.
837 // TODO Should we do the communication in debug mode to support
838 // having an assertion?
839 const bool useDomainDecomposition = (PAR(cr) && !(EI_TPI(inputrec->eI) || inputrec->eI == eiNM));
841 // Note that these variables describe only their own node.
843 // Note that when bonded interactions run on a GPU they always run
844 // alongside a nonbonded task, so do not influence task assignment
845 // even though they affect the force calculation workload.
846 bool useGpuForNonbonded = false;
847 bool useGpuForPme = false;
848 bool useGpuForBonded = false;
849 bool useGpuForUpdate = false;
850 bool gpusWereDetected = hwinfo->ngpu_compatible_tot > 0;
853 // It's possible that there are different numbers of GPUs on
854 // different nodes, which is the user's responsibility to
855 // handle. If unsuitable, we will notice that during task
857 auto canUseGpuForNonbonded = buildSupportsNonbondedOnGpu(nullptr);
858 useGpuForNonbonded = decideWhetherToUseGpusForNonbonded(
859 nonbondedTarget, userGpuTaskAssignment, emulateGpuNonbonded, canUseGpuForNonbonded,
860 gpuAccelerationOfNonbondedIsUseful(mdlog, *inputrec, !GMX_THREAD_MPI), gpusWereDetected);
861 useGpuForPme = decideWhetherToUseGpusForPme(
862 useGpuForNonbonded, pmeTarget, userGpuTaskAssignment, *hwinfo, *inputrec, mtop,
863 cr->nnodes, domdecOptions.numPmeRanks, gpusWereDetected);
864 auto canUseGpuForBonded = buildSupportsGpuBondeds(nullptr)
865 && inputSupportsGpuBondeds(*inputrec, mtop, nullptr);
866 useGpuForBonded = decideWhetherToUseGpusForBonded(
867 useGpuForNonbonded, useGpuForPme, bondedTarget, canUseGpuForBonded,
868 EVDW_PME(inputrec->vdwtype), EEL_PME_EWALD(inputrec->coulombtype),
869 domdecOptions.numPmeRanks, gpusWereDetected);
871 pmeRunMode = (useGpuForPme ? PmeRunMode::GPU : PmeRunMode::CPU);
872 if (pmeRunMode == PmeRunMode::GPU)
874 if (pmeFftTarget == TaskTarget::Cpu)
876 pmeRunMode = PmeRunMode::Mixed;
879 else if (pmeFftTarget == TaskTarget::Gpu)
882 "Assigning FFTs to GPU requires PME to be assigned to GPU as well. With PME "
883 "on CPU you should not be using -pmefft.");
886 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
888 // Initialize development feature flags that enabled by environment variable
889 // and report those features that are enabled.
890 const DevelopmentFeatureFlags devFlags =
891 manageDevelopmentFeatures(mdlog, useGpuForNonbonded, pmeRunMode);
893 // NOTE: The devFlags need decideWhetherToUseGpusForNonbonded(...) and decideWhetherToUseGpusForPme(...) for overrides,
894 // decideWhetherToUseGpuForUpdate() needs devFlags for the '-update auto' override, hence the interleaving.
895 // NOTE: When the simulationWork is constructed, the useGpuForUpdate overrides the devFlags.enableGpuBufferOps.
898 useGpuForUpdate = decideWhetherToUseGpuForUpdate(
899 useDomainDecomposition, useGpuForPme, useGpuForNonbonded, updateTarget,
900 gpusWereDetected, *inputrec, mtop, doEssentialDynamics,
901 gmx_mtop_ftype_count(mtop, F_ORIRES) > 0, replExParams.exchangeInterval > 0);
903 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
907 // TODO: hide restraint implementation details from Mdrunner.
908 // There is nothing unique about restraints at this point as far as the
909 // Mdrunner is concerned. The Mdrunner should just be getting a sequence of
910 // factory functions from the SimulationContext on which to call mdModules_->add().
911 // TODO: capture all restraints into a single RestraintModule, passed to the runner builder.
912 for (auto&& restraint : restraintManager_->getRestraints())
914 auto module = RestraintMDModule::create(restraint, restraint->sites());
915 mdModules_->add(std::move(module));
918 // TODO: Error handling
919 mdModules_->assignOptionsToModules(*inputrec->params, nullptr);
920 const auto& mdModulesNotifier = mdModules_->notifier().notifier_;
922 if (inputrec->internalParameters != nullptr)
924 mdModulesNotifier.notify(*inputrec->internalParameters);
927 if (fplog != nullptr)
929 pr_inputrec(fplog, 0, "Input Parameters", inputrec, FALSE);
930 fprintf(fplog, "\n");
935 /* In rerun, set velocities to zero if present */
936 if (doRerun && ((globalState->flags & (1 << estV)) != 0))
938 // rerun does not use velocities
942 "Rerun trajectory contains velocities. Rerun does only evaluate "
943 "potential energy and forces. The velocities will be ignored.");
944 for (int i = 0; i < globalState->natoms; i++)
946 clear_rvec(globalState->v[i]);
948 globalState->flags &= ~(1 << estV);
951 /* now make sure the state is initialized and propagated */
952 set_state_entries(globalState.get(), inputrec);
955 /* NM and TPI parallelize over force/energy calculations, not atoms,
956 * so we need to initialize and broadcast the global state.
958 if (inputrec->eI == eiNM || inputrec->eI == eiTPI)
962 globalState = std::make_unique<t_state>();
964 broadcastStateWithoutDynamics(cr, globalState.get());
967 /* A parallel command line option consistency check that we can
968 only do after any threads have started. */
970 && (domdecOptions.numCells[XX] > 1 || domdecOptions.numCells[YY] > 1
971 || domdecOptions.numCells[ZZ] > 1 || domdecOptions.numPmeRanks > 0))
974 "The -dd or -npme option request a parallel simulation, "
976 "but %s was compiled without threads or MPI enabled",
977 output_env_get_program_display_name(oenv));
979 "but the number of MPI-threads (option -ntmpi) is not set or is 1");
981 "but %s was not started through mpirun/mpiexec or only one rank was requested "
982 "through mpirun/mpiexec",
983 output_env_get_program_display_name(oenv));
987 if (doRerun && (EI_ENERGY_MINIMIZATION(inputrec->eI) || eiNM == inputrec->eI))
990 "The .mdp file specified an energy mininization or normal mode algorithm, and "
991 "these are not compatible with mdrun -rerun");
994 if (!(EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype)))
996 if (domdecOptions.numPmeRanks > 0)
998 gmx_fatal_collective(FARGS, cr->mpi_comm_mysim, MASTER(cr),
999 "PME-only ranks are requested, but the system does not use PME "
1000 "for electrostatics or LJ");
1003 domdecOptions.numPmeRanks = 0;
1006 if (useGpuForNonbonded && domdecOptions.numPmeRanks < 0)
1008 /* With NB GPUs we don't automatically use PME-only CPU ranks. PME ranks can
1009 * improve performance with many threads per GPU, since our OpenMP
1010 * scaling is bad, but it's difficult to automate the setup.
1012 domdecOptions.numPmeRanks = 0;
1016 if (domdecOptions.numPmeRanks < 0)
1018 domdecOptions.numPmeRanks = 0;
1019 // TODO possibly print a note that one can opt-in for a separate PME GPU rank?
1023 GMX_RELEASE_ASSERT(domdecOptions.numPmeRanks <= 1,
1024 "PME GPU decomposition is not supported");
1031 fcRegisterSteps(inputrec->nsteps, inputrec->init_step);
1035 /* NMR restraints must be initialized before load_checkpoint,
1036 * since with time averaging the history is added to t_state.
1037 * For proper consistency check we therefore need to extend
1039 * So the PME-only nodes (if present) will also initialize
1040 * the distance restraints.
1044 /* This needs to be called before read_checkpoint to extend the state */
1045 init_disres(fplog, &mtop, inputrec, cr, ms, fcd, globalState.get(), replExParams.exchangeInterval > 0);
1047 init_orires(fplog, &mtop, inputrec, cr, ms, globalState.get(), &(fcd->orires));
1049 auto deform = prepareBoxDeformation(globalState->box, cr, *inputrec);
1051 ObservablesHistory observablesHistory = {};
1053 if (startingBehavior != StartingBehavior::NewSimulation)
1055 /* Check if checkpoint file exists before doing continuation.
1056 * This way we can use identical input options for the first and subsequent runs...
1058 if (mdrunOptions.numStepsCommandline > -2)
1060 /* Temporarily set the number of steps to unlimited to avoid
1061 * triggering the nsteps check in load_checkpoint().
1062 * This hack will go away soon when the -nsteps option is removed.
1064 inputrec->nsteps = -1;
1067 load_checkpoint(opt2fn_master("-cpi", filenames.size(), filenames.data(), cr),
1068 logFileHandle, cr, domdecOptions.numCells, inputrec, globalState.get(),
1069 &observablesHistory, mdrunOptions.reproducible, mdModules_->notifier());
1071 if (startingBehavior == StartingBehavior::RestartWithAppending && logFileHandle)
1073 // Now we can start normal logging to the truncated log file.
1074 fplog = gmx_fio_getfp(logFileHandle);
1075 prepareLogAppending(fplog);
1076 logOwner = buildLogger(fplog, MASTER(cr));
1077 mdlog = logOwner.logger();
1081 if (mdrunOptions.numStepsCommandline > -2)
1086 "The -nsteps functionality is deprecated, and may be removed in a future "
1088 "Consider using gmx convert-tpr -nsteps or changing the appropriate .mdp "
1091 /* override nsteps with value set on the commandline */
1092 override_nsteps_cmdline(mdlog, mdrunOptions.numStepsCommandline, inputrec);
1096 copy_mat(globalState->box, box);
1101 gmx_bcast(sizeof(box), box, cr);
1104 if (inputrec->cutoff_scheme != ecutsVERLET)
1107 "This group-scheme .tpr file can no longer be run by mdrun. Please update to the "
1108 "Verlet scheme, or use an earlier version of GROMACS if necessary.");
1110 /* Update rlist and nstlist. */
1111 prepare_verlet_scheme(fplog, cr, inputrec, nstlist_cmdline, &mtop, box,
1112 useGpuForNonbonded || (emulateGpuNonbonded == EmulateGpuNonbonded::Yes),
1115 const bool prefer1DAnd1PulseDD = (devFlags.enableGpuHaloExchange && useGpuForNonbonded);
1116 // This builder is necessary while we have multi-part construction
1117 // of DD. Before DD is constructed, we use the existence of
1118 // the builder object to indicate that further construction of DD
1120 std::unique_ptr<DomainDecompositionBuilder> ddBuilder;
1121 if (useDomainDecomposition)
1123 ddBuilder = std::make_unique<DomainDecompositionBuilder>(
1124 mdlog, cr, domdecOptions, mdrunOptions, prefer1DAnd1PulseDD, mtop, *inputrec, box,
1125 positionsFromStatePointer(globalState.get()));
1129 /* PME, if used, is done on all nodes with 1D decomposition */
1131 cr->duty = (DUTY_PP | DUTY_PME);
1133 if (inputrec->ePBC == epbcSCREW)
1135 gmx_fatal(FARGS, "pbc=screw is only implemented with domain decomposition");
1139 // Produce the task assignment for this rank.
1140 GpuTaskAssignmentsBuilder gpuTaskAssignmentsBuilder;
1141 GpuTaskAssignments gpuTaskAssignments = gpuTaskAssignmentsBuilder.build(
1142 gpuIdsToUse, userGpuTaskAssignment, *hwinfo, cr, ms, physicalNodeComm, nonbondedTarget,
1143 pmeTarget, bondedTarget, updateTarget, useGpuForNonbonded, useGpuForPme,
1144 thisRankHasDuty(cr, DUTY_PP),
1145 // TODO cr->duty & DUTY_PME should imply that a PME
1146 // algorithm is active, but currently does not.
1147 EEL_PME(inputrec->coulombtype) && thisRankHasDuty(cr, DUTY_PME));
1149 const bool printHostName = (cr->nnodes > 1);
1150 gpuTaskAssignments.reportGpuUsage(mdlog, printHostName, useGpuForBonded, pmeRunMode);
1152 // If the user chose a task assignment, give them some hints
1153 // where appropriate.
1154 if (!userGpuTaskAssignment.empty())
1156 gpuTaskAssignments.logPerformanceHints(mdlog, ssize(gpuIdsToUse));
1159 // Get the device handles for the modules, nullptr when no task is assigned.
1160 gmx_device_info_t* nonbondedDeviceInfo = gpuTaskAssignments.initNonbondedDevice(cr);
1161 gmx_device_info_t* pmeDeviceInfo = gpuTaskAssignments.initPmeDevice();
1163 // TODO Initialize GPU streams here.
1165 // TODO Currently this is always built, yet DD partition code
1166 // checks if it is built before using it. Probably it should
1167 // become an MDModule that is made only when another module
1168 // requires it (e.g. pull, CompEl, density fitting), so that we
1169 // don't update the local atom sets unilaterally every step.
1170 LocalAtomSetManager atomSets;
1173 // TODO Pass the GPU streams to ddBuilder to use in buffer
1174 // transfers (e.g. halo exchange)
1175 cr->dd = ddBuilder->build(&atomSets);
1176 // The builder's job is done, so destruct it
1177 ddBuilder.reset(nullptr);
1178 // Note that local state still does not exist yet.
1183 /* After possible communicator splitting in make_dd_communicators.
1184 * we can set up the intra/inter node communication.
1186 gmx_setup_nodecomm(fplog, cr);
1192 GMX_LOG(mdlog.warning)
1194 .appendTextFormatted(
1195 "This is simulation %d out of %d running as a composite GROMACS\n"
1196 "multi-simulation job. Setup for this simulation:\n",
1199 GMX_LOG(mdlog.warning)
1200 .appendTextFormatted("Using %d MPI %s\n", cr->nnodes,
1202 cr->nnodes == 1 ? "thread" : "threads"
1204 cr->nnodes == 1 ? "process" : "processes"
1210 // If mdrun -pin auto honors any affinity setting that already
1211 // exists. If so, it is nice to provide feedback about whether
1212 // that existing affinity setting was from OpenMP or something
1213 // else, so we run this code both before and after we initialize
1214 // the OpenMP support.
1215 gmx_check_thread_affinity_set(mdlog, &hw_opt, hwinfo->nthreads_hw_avail, FALSE);
1216 /* Check and update the number of OpenMP threads requested */
1217 checkAndUpdateRequestedNumOpenmpThreads(&hw_opt, *hwinfo, cr, ms, physicalNodeComm.size_,
1218 pmeRunMode, mtop, *inputrec);
1220 gmx_omp_nthreads_init(mdlog, cr, hwinfo->nthreads_hw_avail, physicalNodeComm.size_,
1221 hw_opt.nthreads_omp, hw_opt.nthreads_omp_pme, !thisRankHasDuty(cr, DUTY_PP));
1223 // Enable FP exception detection, but not in
1224 // Release mode and not for compilers with known buggy FP
1225 // exception support (clang with any optimization) or suspected
1226 // buggy FP exception support (gcc 7.* with optimization).
1227 #if !defined NDEBUG \
1228 && !((defined __clang__ || (defined(__GNUC__) && !defined(__ICC) && __GNUC__ == 7)) \
1229 && defined __OPTIMIZE__)
1230 const bool bEnableFPE = true;
1232 const bool bEnableFPE = false;
1234 // FIXME - reconcile with gmx_feenableexcept() call from CommandLineModuleManager::run()
1237 gmx_feenableexcept();
1240 /* Now that we know the setup is consistent, check for efficiency */
1241 check_resource_division_efficiency(hwinfo, gpuTaskAssignments.thisRankHasAnyGpuTask(),
1242 mdrunOptions.ntompOptionIsSet, cr, mdlog);
1244 /* getting number of PP/PME threads on this MPI / tMPI rank.
1245 PME: env variable should be read only on one node to make sure it is
1246 identical everywhere;
1248 const int numThreadsOnThisRank = thisRankHasDuty(cr, DUTY_PP) ? gmx_omp_nthreads_get(emntNonbonded)
1249 : gmx_omp_nthreads_get(emntPME);
1250 checkHardwareOversubscription(numThreadsOnThisRank, cr->nodeid, *hwinfo->hardwareTopology,
1251 physicalNodeComm, mdlog);
1253 // Enable Peer access between GPUs where available
1254 // Only for DD, only master PP rank needs to perform setup, and only if thread MPI plus
1255 // any of the GPU communication features are active.
1256 if (DOMAINDECOMP(cr) && MASTER(cr) && thisRankHasDuty(cr, DUTY_PP) && GMX_THREAD_MPI
1257 && (devFlags.enableGpuHaloExchange || devFlags.enableGpuPmePPComm))
1259 setupGpuDevicePeerAccess(gpuIdsToUse, mdlog);
1262 if (hw_opt.threadAffinity != ThreadAffinity::Off)
1264 /* Before setting affinity, check whether the affinity has changed
1265 * - which indicates that probably the OpenMP library has changed it
1266 * since we first checked).
1268 gmx_check_thread_affinity_set(mdlog, &hw_opt, hwinfo->nthreads_hw_avail, TRUE);
1270 int numThreadsOnThisNode, intraNodeThreadOffset;
1271 analyzeThreadsOnThisNode(physicalNodeComm, numThreadsOnThisRank, &numThreadsOnThisNode,
1272 &intraNodeThreadOffset);
1274 /* Set the CPU affinity */
1275 gmx_set_thread_affinity(mdlog, cr, &hw_opt, *hwinfo->hardwareTopology, numThreadsOnThisRank,
1276 numThreadsOnThisNode, intraNodeThreadOffset, nullptr);
1279 if (mdrunOptions.timingOptions.resetStep > -1)
1284 "The -resetstep functionality is deprecated, and may be removed in a "
1287 wcycle = wallcycle_init(fplog, mdrunOptions.timingOptions.resetStep, cr);
1291 /* Master synchronizes its value of reset_counters with all nodes
1292 * including PME only nodes */
1293 int64_t reset_counters = wcycle_get_reset_counters(wcycle);
1294 gmx_bcast_sim(sizeof(reset_counters), &reset_counters, cr);
1295 wcycle_set_reset_counters(wcycle, reset_counters);
1298 // Membrane embedding must be initialized before we call init_forcerec()
1303 fprintf(stderr, "Initializing membed");
1305 /* Note that membed cannot work in parallel because mtop is
1306 * changed here. Fix this if we ever want to make it run with
1307 * multiple ranks. */
1308 membed = init_membed(fplog, filenames.size(), filenames.data(), &mtop, inputrec,
1309 globalState.get(), cr, &mdrunOptions.checkpointOptions.period);
1312 const bool thisRankHasPmeGpuTask = gpuTaskAssignments.thisRankHasPmeGpuTask();
1313 std::unique_ptr<MDAtoms> mdAtoms;
1314 std::unique_ptr<gmx_vsite_t> vsite;
1317 if (thisRankHasDuty(cr, DUTY_PP))
1319 mdModulesNotifier.notify(*cr);
1320 mdModulesNotifier.notify(&atomSets);
1321 mdModulesNotifier.notify(PeriodicBoundaryConditionType{ inputrec->ePBC });
1322 mdModulesNotifier.notify(SimulationTimeStep{ inputrec->delta_t });
1323 /* Initiate forcerecord */
1324 fr = new t_forcerec;
1325 fr->forceProviders = mdModules_->initForceProviders();
1326 init_forcerec(fplog, mdlog, fr, fcd, inputrec, &mtop, cr, box,
1327 opt2fn("-table", filenames.size(), filenames.data()),
1328 opt2fn("-tablep", filenames.size(), filenames.data()),
1329 opt2fns("-tableb", filenames.size(), filenames.data()), *hwinfo,
1330 nonbondedDeviceInfo, useGpuForBonded,
1331 pmeRunMode == PmeRunMode::GPU && !thisRankHasDuty(cr, DUTY_PME), pforce, wcycle);
1333 // TODO Move this to happen during domain decomposition setup,
1334 // once stream and event handling works well with that.
1335 // TODO remove need to pass local stream into GPU halo exchange - Redmine #3093
1336 if (havePPDomainDecomposition(cr) && prefer1DAnd1PulseDD && is1DAnd1PulseDD(*cr->dd))
1338 GMX_RELEASE_ASSERT(devFlags.enableGpuBufferOps,
1339 "Must use GMX_USE_GPU_BUFFER_OPS=1 to use GMX_GPU_DD_COMMS=1");
1341 Nbnxm::gpu_get_command_stream(fr->nbv->gpu_nbv, InteractionLocality::Local);
1342 void* streamNonLocal =
1343 Nbnxm::gpu_get_command_stream(fr->nbv->gpu_nbv, InteractionLocality::NonLocal);
1344 GMX_LOG(mdlog.warning)
1346 .appendTextFormatted(
1347 "NOTE: This run uses the 'GPU halo exchange' feature, enabled by the "
1348 "GMX_GPU_DD_COMMS environment variable.");
1349 cr->dd->gpuHaloExchange = std::make_unique<GpuHaloExchange>(
1350 cr->dd, cr->mpi_comm_mysim, streamLocal, streamNonLocal);
1353 /* Initialize the mdAtoms structure.
1354 * mdAtoms is not filled with atom data,
1355 * as this can not be done now with domain decomposition.
1357 mdAtoms = makeMDAtoms(fplog, mtop, *inputrec, thisRankHasPmeGpuTask);
1358 if (globalState && thisRankHasPmeGpuTask)
1360 // The pinning of coordinates in the global state object works, because we only use
1361 // PME on GPU without DD or on a separate PME rank, and because the local state pointer
1362 // points to the global state object without DD.
1363 // FIXME: MD and EM separately set up the local state - this should happen in the same
1364 // function, which should also perform the pinning.
1365 changePinningPolicy(&globalState->x, pme_get_pinning_policy());
1368 /* Initialize the virtual site communication */
1369 vsite = initVsite(mtop, cr);
1371 calc_shifts(box, fr->shift_vec);
1373 /* With periodic molecules the charge groups should be whole at start up
1374 * and the virtual sites should not be far from their proper positions.
1376 if (!inputrec->bContinuation && MASTER(cr) && !(inputrec->ePBC != epbcNONE && inputrec->bPeriodicMols))
1378 /* Make molecules whole at start of run */
1379 if (fr->ePBC != epbcNONE)
1381 do_pbc_first_mtop(fplog, inputrec->ePBC, box, &mtop, globalState->x.rvec_array());
1385 /* Correct initial vsite positions are required
1386 * for the initial distribution in the domain decomposition
1387 * and for the initial shell prediction.
1389 constructVsitesGlobal(mtop, globalState->x);
1393 if (EEL_PME(fr->ic->eeltype) || EVDW_PME(fr->ic->vdwtype))
1395 ewaldcoeff_q = fr->ic->ewaldcoeff_q;
1396 ewaldcoeff_lj = fr->ic->ewaldcoeff_lj;
1401 /* This is a PME only node */
1403 GMX_ASSERT(globalState == nullptr,
1404 "We don't need the state on a PME only rank and expect it to be unitialized");
1406 ewaldcoeff_q = calc_ewaldcoeff_q(inputrec->rcoulomb, inputrec->ewald_rtol);
1407 ewaldcoeff_lj = calc_ewaldcoeff_lj(inputrec->rvdw, inputrec->ewald_rtol_lj);
1410 gmx_pme_t* sepPmeData = nullptr;
1411 // This reference hides the fact that PME data is owned by runner on PME-only ranks and by forcerec on other ranks
1412 GMX_ASSERT(thisRankHasDuty(cr, DUTY_PP) == (fr != nullptr),
1413 "Double-checking that only PME-only ranks have no forcerec");
1414 gmx_pme_t*& pmedata = fr ? fr->pmedata : sepPmeData;
1416 // TODO should live in ewald module once its testing is improved
1418 // Later, this program could contain kernels that might be later
1419 // re-used as auto-tuning progresses, or subsequent simulations
1421 PmeGpuProgramStorage pmeGpuProgram;
1422 if (thisRankHasPmeGpuTask)
1424 pmeGpuProgram = buildPmeGpuProgram(pmeDeviceInfo);
1427 /* Initiate PME if necessary,
1428 * either on all nodes or on dedicated PME nodes only. */
1429 if (EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype))
1431 if (mdAtoms && mdAtoms->mdatoms())
1433 nChargePerturbed = mdAtoms->mdatoms()->nChargePerturbed;
1434 if (EVDW_PME(inputrec->vdwtype))
1436 nTypePerturbed = mdAtoms->mdatoms()->nTypePerturbed;
1439 if (cr->npmenodes > 0)
1441 /* The PME only nodes need to know nChargePerturbed(FEP on Q) and nTypePerturbed(FEP on LJ)*/
1442 gmx_bcast_sim(sizeof(nChargePerturbed), &nChargePerturbed, cr);
1443 gmx_bcast_sim(sizeof(nTypePerturbed), &nTypePerturbed, cr);
1446 if (thisRankHasDuty(cr, DUTY_PME))
1450 pmedata = gmx_pme_init(cr, getNumPmeDomains(cr->dd), inputrec, nChargePerturbed != 0,
1451 nTypePerturbed != 0, mdrunOptions.reproducible, ewaldcoeff_q,
1452 ewaldcoeff_lj, gmx_omp_nthreads_get(emntPME), pmeRunMode,
1453 nullptr, pmeDeviceInfo, pmeGpuProgram.get(), mdlog);
1455 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1460 if (EI_DYNAMICS(inputrec->eI))
1462 /* Turn on signal handling on all nodes */
1464 * (A user signal from the PME nodes (if any)
1465 * is communicated to the PP nodes.
1467 signal_handler_install();
1470 pull_t* pull_work = nullptr;
1471 if (thisRankHasDuty(cr, DUTY_PP))
1473 /* Assumes uniform use of the number of OpenMP threads */
1474 walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(emntDefault));
1476 if (inputrec->bPull)
1478 /* Initialize pull code */
1479 pull_work = init_pull(fplog, inputrec->pull, inputrec, &mtop, cr, &atomSets,
1480 inputrec->fepvals->init_lambda);
1481 if (inputrec->pull->bXOutAverage || inputrec->pull->bFOutAverage)
1483 initPullHistory(pull_work, &observablesHistory);
1485 if (EI_DYNAMICS(inputrec->eI) && MASTER(cr))
1487 init_pull_output_files(pull_work, filenames.size(), filenames.data(), oenv, startingBehavior);
1491 std::unique_ptr<EnforcedRotation> enforcedRotation;
1494 /* Initialize enforced rotation code */
1496 init_rot(fplog, inputrec, filenames.size(), filenames.data(), cr, &atomSets,
1497 globalState.get(), &mtop, oenv, mdrunOptions, startingBehavior);
1500 t_swap* swap = nullptr;
1501 if (inputrec->eSwapCoords != eswapNO)
1503 /* Initialize ion swapping code */
1504 swap = init_swapcoords(fplog, inputrec,
1505 opt2fn_master("-swap", filenames.size(), filenames.data(), cr),
1506 &mtop, globalState.get(), &observablesHistory, cr, &atomSets,
1507 oenv, mdrunOptions, startingBehavior);
1510 /* Let makeConstraints know whether we have essential dynamics constraints. */
1511 auto constr = makeConstraints(mtop, *inputrec, pull_work, doEssentialDynamics, fplog,
1512 *mdAtoms->mdatoms(), cr, ms, &nrnb, wcycle, fr->bMolPBC);
1514 /* Energy terms and groups */
1515 gmx_enerdata_t enerd(mtop.groups.groups[SimulationAtomGroupType::EnergyOutput].size(),
1516 inputrec->fepvals->n_lambda);
1518 /* Kinetic energy data */
1519 gmx_ekindata_t ekind;
1520 init_ekindata(fplog, &mtop, &(inputrec->opts), &ekind, inputrec->cos_accel);
1522 /* Set up interactive MD (IMD) */
1524 makeImdSession(inputrec, cr, wcycle, &enerd, ms, &mtop, mdlog,
1525 MASTER(cr) ? globalState->x.rvec_array() : nullptr, filenames.size(),
1526 filenames.data(), oenv, mdrunOptions.imdOptions, startingBehavior);
1528 if (DOMAINDECOMP(cr))
1530 GMX_RELEASE_ASSERT(fr, "fr was NULL while cr->duty was DUTY_PP");
1531 /* This call is not included in init_domain_decomposition mainly
1532 * because fr->cginfo_mb is set later.
1534 dd_init_bondeds(fplog, cr->dd, &mtop, vsite.get(), inputrec,
1535 domdecOptions.checkBondedInteractions, fr->cginfo_mb);
1538 const bool inputIsCompatibleWithModularSimulator = ModularSimulator::isInputCompatible(
1539 false, inputrec, doRerun, vsite.get(), ms, replExParams, fcd,
1540 static_cast<int>(filenames.size()), filenames.data(), &observablesHistory, membed);
1542 const bool useModularSimulator = inputIsCompatibleWithModularSimulator
1543 && !(getenv("GMX_DISABLE_MODULAR_SIMULATOR") != nullptr);
1545 // TODO This is not the right place to manage the lifetime of
1546 // this data structure, but currently it's the easiest way to
1548 MdrunScheduleWorkload runScheduleWork;
1549 // Also populates the simulation constant workload description.
1550 runScheduleWork.simulationWork = createSimulationWorkload(
1551 useGpuForNonbonded, pmeRunMode, useGpuForBonded, useGpuForUpdate,
1552 devFlags.enableGpuBufferOps, devFlags.enableGpuHaloExchange,
1553 devFlags.enableGpuPmePPComm, haveEwaldSurfaceContribution(*inputrec));
1555 std::unique_ptr<gmx::StatePropagatorDataGpu> stateGpu;
1556 if (gpusWereDetected
1557 && ((useGpuForPme && thisRankHasDuty(cr, DUTY_PME))
1558 || runScheduleWork.simulationWork.useGpuBufferOps))
1560 const void* pmeStream = pme_gpu_get_device_stream(fr->pmedata);
1561 const void* localStream =
1562 fr->nbv->gpu_nbv != nullptr
1563 ? Nbnxm::gpu_get_command_stream(fr->nbv->gpu_nbv, InteractionLocality::Local)
1565 const void* nonLocalStream =
1566 fr->nbv->gpu_nbv != nullptr
1567 ? Nbnxm::gpu_get_command_stream(fr->nbv->gpu_nbv, InteractionLocality::NonLocal)
1569 const void* deviceContext = pme_gpu_get_device_context(fr->pmedata);
1570 const int paddingSize = pme_gpu_get_padding_size(fr->pmedata);
1571 GpuApiCallBehavior transferKind = (inputrec->eI == eiMD && !doRerun && !useModularSimulator)
1572 ? GpuApiCallBehavior::Async
1573 : GpuApiCallBehavior::Sync;
1575 stateGpu = std::make_unique<gmx::StatePropagatorDataGpu>(
1576 pmeStream, localStream, nonLocalStream, deviceContext, transferKind, paddingSize);
1577 fr->stateGpu = stateGpu.get();
1580 GMX_ASSERT(stopHandlerBuilder_, "Runner must provide StopHandlerBuilder to simulator.");
1581 SimulatorBuilder simulatorBuilder;
1583 // build and run simulator object based on user-input
1584 auto simulator = simulatorBuilder.build(
1585 inputIsCompatibleWithModularSimulator, fplog, cr, ms, mdlog,
1586 static_cast<int>(filenames.size()), filenames.data(), oenv, mdrunOptions,
1587 startingBehavior, vsite.get(), constr.get(),
1588 enforcedRotation ? enforcedRotation->getLegacyEnfrot() : nullptr, deform.get(),
1589 mdModules_->outputProvider(), mdModules_->notifier(), inputrec, imdSession.get(),
1590 pull_work, swap, &mtop, fcd, globalState.get(), &observablesHistory, mdAtoms.get(),
1591 &nrnb, wcycle, fr, &enerd, &ekind, &runScheduleWork, replExParams, membed,
1592 walltime_accounting, std::move(stopHandlerBuilder_), doRerun);
1595 if (fr->pmePpCommGpu)
1597 // destroy object since it is no longer required. (This needs to be done while the GPU context still exists.)
1598 fr->pmePpCommGpu.reset();
1601 if (inputrec->bPull)
1603 finish_pull(pull_work);
1605 finish_swapcoords(swap);
1609 GMX_RELEASE_ASSERT(pmedata, "pmedata was NULL while cr->duty was not DUTY_PP");
1611 walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(emntPME));
1612 gmx_pmeonly(pmedata, cr, &nrnb, wcycle, walltime_accounting, inputrec, pmeRunMode);
1615 wallcycle_stop(wcycle, ewcRUN);
1617 /* Finish up, write some stuff
1618 * if rerunMD, don't write last frame again
1620 finish_run(fplog, mdlog, cr, inputrec, &nrnb, wcycle, walltime_accounting,
1621 fr ? fr->nbv.get() : nullptr, pmedata, EI_DYNAMICS(inputrec->eI) && !isMultiSim(ms));
1623 // clean up cycle counter
1624 wallcycle_destroy(wcycle);
1629 gmx_pme_destroy(pmedata);
1633 // FIXME: this is only here to manually unpin mdAtoms->chargeA_ and state->x,
1634 // before we destroy the GPU context(s) in free_gpu_resources().
1635 // Pinned buffers are associated with contexts in CUDA.
1636 // As soon as we destroy GPU contexts after mdrunner() exits, these lines should go.
1637 mdAtoms.reset(nullptr);
1638 globalState.reset(nullptr);
1639 mdModules_.reset(nullptr); // destruct force providers here as they might also use the GPU
1641 /* Free GPU memory and set a physical node tMPI barrier (which should eventually go away) */
1642 free_gpu_resources(fr, physicalNodeComm, hwinfo->gpu_info);
1643 free_gpu(nonbondedDeviceInfo);
1644 free_gpu(pmeDeviceInfo);
1645 done_forcerec(fr, mtop.molblock.size());
1650 free_membed(membed);
1653 /* Does what it says */
1654 print_date_and_time(fplog, cr->nodeid, "Finished mdrun", gmx_gettime());
1655 walltime_accounting_destroy(walltime_accounting);
1657 // Ensure log file content is written
1660 gmx_fio_flush(logFileHandle);
1663 /* Reset FPEs (important for unit tests) by disabling them. Assumes no
1664 * exceptions were enabled before function was called. */
1667 gmx_fedisableexcept();
1670 auto rc = static_cast<int>(gmx_get_stop_condition());
1673 /* we need to join all threads. The sub-threads join when they
1674 exit this function, but the master thread needs to be told to
1676 if (PAR(cr) && MASTER(cr))
1684 Mdrunner::~Mdrunner()
1686 // Clean up of the Manager.
1687 // This will end up getting called on every thread-MPI rank, which is unnecessary,
1688 // but okay as long as threads synchronize some time before adding or accessing
1689 // a new set of restraints.
1690 if (restraintManager_)
1692 restraintManager_->clear();
1693 GMX_ASSERT(restraintManager_->countRestraints() == 0,
1694 "restraints added during runner life time should be cleared at runner "
1699 void Mdrunner::addPotential(std::shared_ptr<gmx::IRestraintPotential> puller, const std::string& name)
1701 GMX_ASSERT(restraintManager_, "Mdrunner must have a restraint manager.");
1702 // Not sure if this should be logged through the md logger or something else,
1703 // but it is helpful to have some sort of INFO level message sent somewhere.
1704 // std::cout << "Registering restraint named " << name << std::endl;
1706 // When multiple restraints are used, it may be wasteful to register them separately.
1707 // Maybe instead register an entire Restraint Manager as a force provider.
1708 restraintManager_->addToSpec(std::move(puller), name);
1711 Mdrunner::Mdrunner(std::unique_ptr<MDModules> mdModules) : mdModules_(std::move(mdModules)) {}
1713 Mdrunner::Mdrunner(Mdrunner&&) noexcept = default;
1715 //NOLINTNEXTLINE(performance-noexcept-move-constructor) working around GCC bug 58265
1716 Mdrunner& Mdrunner::operator=(Mdrunner&& /*handle*/) noexcept(BUGFREE_NOEXCEPT_STRING) = default;
1718 class Mdrunner::BuilderImplementation
1721 BuilderImplementation() = delete;
1722 BuilderImplementation(std::unique_ptr<MDModules> mdModules, compat::not_null<SimulationContext*> context);
1723 ~BuilderImplementation();
1725 BuilderImplementation& setExtraMdrunOptions(const MdrunOptions& options,
1726 real forceWarningThreshold,
1727 StartingBehavior startingBehavior);
1729 void addDomdec(const DomdecOptions& options);
1731 void addVerletList(int nstlist);
1733 void addReplicaExchange(const ReplicaExchangeParameters& params);
1735 void addNonBonded(const char* nbpu_opt);
1737 void addPME(const char* pme_opt_, const char* pme_fft_opt_);
1739 void addBondedTaskAssignment(const char* bonded_opt);
1741 void addUpdateTaskAssignment(const char* update_opt);
1743 void addHardwareOptions(const gmx_hw_opt_t& hardwareOptions);
1745 void addFilenames(ArrayRef<const t_filenm> filenames);
1747 void addOutputEnvironment(gmx_output_env_t* outputEnvironment);
1749 void addLogFile(t_fileio* logFileHandle);
1751 void addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder);
1756 // Default parameters copied from runner.h
1757 // \todo Clarify source(s) of default parameters.
1759 const char* nbpu_opt_ = nullptr;
1760 const char* pme_opt_ = nullptr;
1761 const char* pme_fft_opt_ = nullptr;
1762 const char* bonded_opt_ = nullptr;
1763 const char* update_opt_ = nullptr;
1765 MdrunOptions mdrunOptions_;
1767 DomdecOptions domdecOptions_;
1769 ReplicaExchangeParameters replicaExchangeParameters_;
1771 //! Command-line override for the duration of a neighbor list with the Verlet scheme.
1774 //! Multisim communicator handle.
1775 gmx_multisim_t* multiSimulation_;
1777 //! mdrun communicator
1778 MPI_Comm communicator_ = MPI_COMM_NULL;
1780 //! Print a warning if any force is larger than this (in kJ/mol nm).
1781 real forceWarningThreshold_ = -1;
1783 //! Whether the simulation will start afresh, or restart with/without appending.
1784 StartingBehavior startingBehavior_ = StartingBehavior::NewSimulation;
1786 //! The modules that comprise the functionality of mdrun.
1787 std::unique_ptr<MDModules> mdModules_;
1789 //! \brief Parallelism information.
1790 gmx_hw_opt_t hardwareOptions_;
1792 //! filename options for simulation.
1793 ArrayRef<const t_filenm> filenames_;
1795 /*! \brief Handle to output environment.
1797 * \todo gmx_output_env_t needs lifetime management.
1799 gmx_output_env_t* outputEnvironment_ = nullptr;
1801 /*! \brief Non-owning handle to MD log file.
1803 * \todo Context should own output facilities for client.
1804 * \todo Improve log file handle management.
1806 * Code managing the FILE* relies on the ability to set it to
1807 * nullptr to check whether the filehandle is valid.
1809 t_fileio* logFileHandle_ = nullptr;
1812 * \brief Builder for simulation stop signal handler.
1814 std::unique_ptr<StopHandlerBuilder> stopHandlerBuilder_ = nullptr;
1817 Mdrunner::BuilderImplementation::BuilderImplementation(std::unique_ptr<MDModules> mdModules,
1818 compat::not_null<SimulationContext*> context) :
1819 mdModules_(std::move(mdModules))
1821 communicator_ = context->communicator_;
1822 multiSimulation_ = context->multiSimulation_.get();
1825 Mdrunner::BuilderImplementation::~BuilderImplementation() = default;
1827 Mdrunner::BuilderImplementation&
1828 Mdrunner::BuilderImplementation::setExtraMdrunOptions(const MdrunOptions& options,
1829 const real forceWarningThreshold,
1830 const StartingBehavior startingBehavior)
1832 mdrunOptions_ = options;
1833 forceWarningThreshold_ = forceWarningThreshold;
1834 startingBehavior_ = startingBehavior;
1838 void Mdrunner::BuilderImplementation::addDomdec(const DomdecOptions& options)
1840 domdecOptions_ = options;
1843 void Mdrunner::BuilderImplementation::addVerletList(int nstlist)
1848 void Mdrunner::BuilderImplementation::addReplicaExchange(const ReplicaExchangeParameters& params)
1850 replicaExchangeParameters_ = params;
1853 Mdrunner Mdrunner::BuilderImplementation::build()
1855 auto newRunner = Mdrunner(std::move(mdModules_));
1857 newRunner.mdrunOptions = mdrunOptions_;
1858 newRunner.pforce = forceWarningThreshold_;
1859 newRunner.startingBehavior = startingBehavior_;
1860 newRunner.domdecOptions = domdecOptions_;
1862 // \todo determine an invariant to check or confirm that all gmx_hw_opt_t objects are valid
1863 newRunner.hw_opt = hardwareOptions_;
1865 // No invariant to check. This parameter exists to optionally override other behavior.
1866 newRunner.nstlist_cmdline = nstlist_;
1868 newRunner.replExParams = replicaExchangeParameters_;
1870 newRunner.filenames = filenames_;
1872 newRunner.communicator = communicator_;
1874 // nullptr is a valid value for the multisim handle
1875 newRunner.ms = multiSimulation_;
1877 // \todo Clarify ownership and lifetime management for gmx_output_env_t
1878 // \todo Update sanity checking when output environment has clearly specified invariants.
1879 // Initialization and default values for oenv are not well specified in the current version.
1880 if (outputEnvironment_)
1882 newRunner.oenv = outputEnvironment_;
1886 GMX_THROW(gmx::APIError(
1887 "MdrunnerBuilder::addOutputEnvironment() is required before build()"));
1890 newRunner.logFileHandle = logFileHandle_;
1894 newRunner.nbpu_opt = nbpu_opt_;
1898 GMX_THROW(gmx::APIError("MdrunnerBuilder::addNonBonded() is required before build()"));
1901 if (pme_opt_ && pme_fft_opt_)
1903 newRunner.pme_opt = pme_opt_;
1904 newRunner.pme_fft_opt = pme_fft_opt_;
1908 GMX_THROW(gmx::APIError("MdrunnerBuilder::addElectrostatics() is required before build()"));
1913 newRunner.bonded_opt = bonded_opt_;
1917 GMX_THROW(gmx::APIError(
1918 "MdrunnerBuilder::addBondedTaskAssignment() is required before build()"));
1923 newRunner.update_opt = update_opt_;
1927 GMX_THROW(gmx::APIError(
1928 "MdrunnerBuilder::addUpdateTaskAssignment() is required before build() "));
1932 newRunner.restraintManager_ = std::make_unique<gmx::RestraintManager>();
1934 if (stopHandlerBuilder_)
1936 newRunner.stopHandlerBuilder_ = std::move(stopHandlerBuilder_);
1940 newRunner.stopHandlerBuilder_ = std::make_unique<StopHandlerBuilder>();
1946 void Mdrunner::BuilderImplementation::addNonBonded(const char* nbpu_opt)
1948 nbpu_opt_ = nbpu_opt;
1951 void Mdrunner::BuilderImplementation::addPME(const char* pme_opt, const char* pme_fft_opt)
1954 pme_fft_opt_ = pme_fft_opt;
1957 void Mdrunner::BuilderImplementation::addBondedTaskAssignment(const char* bonded_opt)
1959 bonded_opt_ = bonded_opt;
1962 void Mdrunner::BuilderImplementation::addUpdateTaskAssignment(const char* update_opt)
1964 update_opt_ = update_opt;
1967 void Mdrunner::BuilderImplementation::addHardwareOptions(const gmx_hw_opt_t& hardwareOptions)
1969 hardwareOptions_ = hardwareOptions;
1972 void Mdrunner::BuilderImplementation::addFilenames(ArrayRef<const t_filenm> filenames)
1974 filenames_ = filenames;
1977 void Mdrunner::BuilderImplementation::addOutputEnvironment(gmx_output_env_t* outputEnvironment)
1979 outputEnvironment_ = outputEnvironment;
1982 void Mdrunner::BuilderImplementation::addLogFile(t_fileio* logFileHandle)
1984 logFileHandle_ = logFileHandle;
1987 void Mdrunner::BuilderImplementation::addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder)
1989 stopHandlerBuilder_ = std::move(builder);
1992 MdrunnerBuilder::MdrunnerBuilder(std::unique_ptr<MDModules> mdModules,
1993 compat::not_null<SimulationContext*> context) :
1994 impl_{ std::make_unique<Mdrunner::BuilderImplementation>(std::move(mdModules), context) }
1998 MdrunnerBuilder::~MdrunnerBuilder() = default;
2000 MdrunnerBuilder& MdrunnerBuilder::addSimulationMethod(const MdrunOptions& options,
2001 real forceWarningThreshold,
2002 const StartingBehavior startingBehavior)
2004 impl_->setExtraMdrunOptions(options, forceWarningThreshold, startingBehavior);
2008 MdrunnerBuilder& MdrunnerBuilder::addDomainDecomposition(const DomdecOptions& options)
2010 impl_->addDomdec(options);
2014 MdrunnerBuilder& MdrunnerBuilder::addNeighborList(int nstlist)
2016 impl_->addVerletList(nstlist);
2020 MdrunnerBuilder& MdrunnerBuilder::addReplicaExchange(const ReplicaExchangeParameters& params)
2022 impl_->addReplicaExchange(params);
2026 MdrunnerBuilder& MdrunnerBuilder::addNonBonded(const char* nbpu_opt)
2028 impl_->addNonBonded(nbpu_opt);
2032 MdrunnerBuilder& MdrunnerBuilder::addElectrostatics(const char* pme_opt, const char* pme_fft_opt)
2034 // The builder method may become more general in the future, but in this version,
2035 // parameters for PME electrostatics are both required and the only parameters
2037 if (pme_opt && pme_fft_opt)
2039 impl_->addPME(pme_opt, pme_fft_opt);
2044 gmx::InvalidInputError("addElectrostatics() arguments must be non-null pointers."));
2049 MdrunnerBuilder& MdrunnerBuilder::addBondedTaskAssignment(const char* bonded_opt)
2051 impl_->addBondedTaskAssignment(bonded_opt);
2055 MdrunnerBuilder& MdrunnerBuilder::addUpdateTaskAssignment(const char* update_opt)
2057 impl_->addUpdateTaskAssignment(update_opt);
2061 Mdrunner MdrunnerBuilder::build()
2063 return impl_->build();
2066 MdrunnerBuilder& MdrunnerBuilder::addHardwareOptions(const gmx_hw_opt_t& hardwareOptions)
2068 impl_->addHardwareOptions(hardwareOptions);
2072 MdrunnerBuilder& MdrunnerBuilder::addFilenames(ArrayRef<const t_filenm> filenames)
2074 impl_->addFilenames(filenames);
2078 MdrunnerBuilder& MdrunnerBuilder::addOutputEnvironment(gmx_output_env_t* outputEnvironment)
2080 impl_->addOutputEnvironment(outputEnvironment);
2084 MdrunnerBuilder& MdrunnerBuilder::addLogFile(t_fileio* logFileHandle)
2086 impl_->addLogFile(logFileHandle);
2090 MdrunnerBuilder& MdrunnerBuilder::addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder)
2092 impl_->addStopHandlerBuilder(std::move(builder));
2096 MdrunnerBuilder::MdrunnerBuilder(MdrunnerBuilder&&) noexcept = default;
2098 MdrunnerBuilder& MdrunnerBuilder::operator=(MdrunnerBuilder&&) noexcept = default;