2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
5 * Copyright (c) 2001-2004, The GROMACS development team.
6 * Copyright (c) 2011,2012,2013,2014,2015,2016,2017,2018,2019, by the GROMACS development team, led by
7 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
8 * and including many others, as listed in the AUTHORS file in the
9 * top-level source directory and at http://www.gromacs.org.
11 * GROMACS is free software; you can redistribute it and/or
12 * modify it under the terms of the GNU Lesser General Public License
13 * as published by the Free Software Foundation; either version 2.1
14 * of the License, or (at your option) any later version.
16 * GROMACS is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
19 * Lesser General Public License for more details.
21 * You should have received a copy of the GNU Lesser General Public
22 * License along with GROMACS; if not, see
23 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
24 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
26 * If you want to redistribute modifications to GROMACS, please
27 * consider that scientific software is very special. Version
28 * control is crucial - bugs must be traceable. We will be happy to
29 * consider code for inclusion in the official distribution, but
30 * derived work must not be called official GROMACS. Details are found
31 * in the README & COPYING files - if they are missing, get the
32 * official version at http://www.gromacs.org.
34 * To help us fund GROMACS development, we humbly ask that you cite
35 * the research papers on the package. Check out http://www.gromacs.org.
39 * \brief Implements the MD runner routine calling all integrators.
41 * \author David van der Spoel <david.vanderspoel@icm.uu.se>
42 * \ingroup module_mdrun
59 #include "gromacs/commandline/filenm.h"
60 #include "gromacs/domdec/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 //! If true, forces 'mdrun -update auto' default to 'gpu'
179 bool forceGpuUpdateDefaultOn = false;
180 //! True if the GPU halo exchange development feature is enabled
181 bool enableGpuHaloExchange = false;
182 //! True if the PME PP direct communication GPU development feature is enabled
183 bool enableGpuPmePPComm = false;
186 /*! \brief Manage any development feature flag variables encountered
188 * The use of dev features indicated by environment variables is
189 * logged in order to ensure that runs with such features enabled can
190 * be identified from their log and standard output. Any cross
191 * dependencies are also checked, and if unsatisfied, a fatal error
194 * Note that some development features overrides are applied already here:
195 * the GPU communication flags are set to false in non-tMPI and non-CUDA builds.
197 * \param[in] mdlog Logger object.
198 * \param[in] useGpuForNonbonded True if the nonbonded task is offloaded in this run.
199 * \param[in] useGpuForPme True if the PME task is offloaded in this run.
200 * \returns The object populated with development feature flags.
202 static DevelopmentFeatureFlags manageDevelopmentFeatures(const gmx::MDLogger &mdlog,
203 const bool useGpuForNonbonded,
204 const bool useGpuForPme)
206 DevelopmentFeatureFlags devFlags;
208 // Some builds of GCC 5 give false positive warnings that these
209 // getenv results are ignored when clearly they are used.
210 #pragma GCC diagnostic push
211 #pragma GCC diagnostic ignored "-Wunused-result"
212 devFlags.enableGpuBufferOps = (getenv("GMX_USE_GPU_BUFFER_OPS") != nullptr) && (GMX_GPU == GMX_GPU_CUDA) && useGpuForNonbonded;
213 devFlags.forceGpuUpdateDefaultOn = (getenv("GMX_FORCE_UPDATE_DEFAULT_GPU") != nullptr);
214 devFlags.enableGpuHaloExchange = (getenv("GMX_GPU_DD_COMMS") != nullptr && GMX_THREAD_MPI && (GMX_GPU == GMX_GPU_CUDA));
215 devFlags.enableGpuPmePPComm = (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).asParagraph().appendTextFormatted(
221 "NOTE: This run uses the 'GPU buffer ops' feature, enabled by the GMX_USE_GPU_BUFFER_OPS environment variable.");
224 if (devFlags.enableGpuHaloExchange)
226 if (useGpuForNonbonded)
228 if (!devFlags.enableGpuBufferOps)
230 gmx_fatal(FARGS, "Cannot enable GPU halo exchange without GPU buffer operations, set GMX_USE_GPU_BUFFER_OPS=1\n");
232 GMX_LOG(mdlog.warning).asParagraph().appendTextFormatted(
233 "NOTE: This run uses the 'GPU halo exchange' feature, enabled by the GMX_GPU_DD_COMMS environment variable.");
237 GMX_LOG(mdlog.warning).asParagraph().appendTextFormatted(
238 "NOTE: GMX_GPU_DD_COMMS environment variable detected, but the 'GPU halo exchange' feature will not be enabled as nonbonded interactions are not offloaded.");
239 devFlags.enableGpuHaloExchange = false;
243 if (devFlags.forceGpuUpdateDefaultOn)
245 GMX_LOG(mdlog.warning).asParagraph().appendTextFormatted(
246 "NOTE: This run will default to '-update gpu' as requested by the GMX_FORCE_UPDATE_DEFAULT_GPU environment variable.");
249 if (devFlags.enableGpuPmePPComm)
253 GMX_LOG(mdlog.warning).asParagraph().appendTextFormatted(
254 "NOTE: This run uses the 'GPU PME-PP communications' feature, enabled by the GMX_GPU_PME_PP_COMMS environment variable.");
258 GMX_LOG(mdlog.warning).asParagraph().appendTextFormatted(
259 "NOTE: GMX_GPU_PME_PP_COMMS environment variable detected, but the 'GPU PME-PP communications' feature was not enabled as PME is not offloaded to the GPU.");
260 devFlags.enableGpuPmePPComm = false;
267 /*! \brief Barrier for safe simultaneous thread access to mdrunner data
269 * Used to ensure that the master thread does not modify mdrunner during copy
270 * on the spawned threads. */
271 static void threadMpiMdrunnerAccessBarrier()
274 MPI_Barrier(MPI_COMM_WORLD);
278 Mdrunner Mdrunner::cloneOnSpawnedThread() const
280 auto newRunner = Mdrunner(std::make_unique<MDModules>());
282 // All runners in the same process share a restraint manager resource because it is
283 // part of the interface to the client code, which is associated only with the
284 // original thread. Handles to the same resources can be obtained by copy.
286 newRunner.restraintManager_ = std::make_unique<RestraintManager>(*restraintManager_);
289 // Copy members of master runner.
290 // \todo Replace with builder when Simulation context and/or runner phases are better defined.
291 // Ref https://redmine.gromacs.org/issues/2587 and https://redmine.gromacs.org/issues/2375
292 newRunner.hw_opt = hw_opt;
293 newRunner.filenames = filenames;
295 newRunner.oenv = oenv;
296 newRunner.mdrunOptions = mdrunOptions;
297 newRunner.domdecOptions = domdecOptions;
298 newRunner.nbpu_opt = nbpu_opt;
299 newRunner.pme_opt = pme_opt;
300 newRunner.pme_fft_opt = pme_fft_opt;
301 newRunner.bonded_opt = bonded_opt;
302 newRunner.update_opt = update_opt;
303 newRunner.nstlist_cmdline = nstlist_cmdline;
304 newRunner.replExParams = replExParams;
305 newRunner.pforce = pforce;
306 // Give the spawned thread the newly created valid communicator
307 // for the simulation.
308 newRunner.communicator = MPI_COMM_WORLD;
310 newRunner.startingBehavior = startingBehavior;
311 newRunner.stopHandlerBuilder_ = std::make_unique<StopHandlerBuilder>(*stopHandlerBuilder_);
313 threadMpiMdrunnerAccessBarrier();
318 /*! \brief The callback used for running on spawned threads.
320 * Obtains the pointer to the master mdrunner object from the one
321 * argument permitted to the thread-launch API call, copies it to make
322 * a new runner for this thread, reinitializes necessary data, and
323 * proceeds to the simulation. */
324 static void mdrunner_start_fn(const void *arg)
328 auto masterMdrunner = reinterpret_cast<const gmx::Mdrunner *>(arg);
329 /* copy the arg list to make sure that it's thread-local. This
330 doesn't copy pointed-to items, of course; fnm, cr and fplog
331 are reset in the call below, all others should be const. */
332 gmx::Mdrunner mdrunner = masterMdrunner->cloneOnSpawnedThread();
335 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
339 void Mdrunner::spawnThreads(int numThreadsToLaunch)
342 /* now spawn new threads that start mdrunner_start_fn(), while
343 the main thread returns. Thread affinity is handled later. */
344 if (tMPI_Init_fn(TRUE, numThreadsToLaunch, TMPI_AFFINITY_NONE,
345 mdrunner_start_fn, static_cast<const void*>(this)) != TMPI_SUCCESS)
347 GMX_THROW(gmx::InternalError("Failed to spawn thread-MPI threads"));
350 // Give the master thread the newly created valid communicator for
352 communicator = MPI_COMM_WORLD;
353 threadMpiMdrunnerAccessBarrier();
355 GMX_UNUSED_VALUE(numThreadsToLaunch);
356 GMX_UNUSED_VALUE(mdrunner_start_fn);
362 /*! \brief Initialize variables for Verlet scheme simulation */
363 static void prepare_verlet_scheme(FILE *fplog,
367 const gmx_mtop_t *mtop,
369 bool makeGpuPairList,
370 const gmx::CpuInfo &cpuinfo)
372 /* For NVE simulations, we will retain the initial list buffer */
373 if (EI_DYNAMICS(ir->eI) &&
374 ir->verletbuf_tol > 0 &&
375 !(EI_MD(ir->eI) && ir->etc == etcNO))
377 /* Update the Verlet buffer size for the current run setup */
379 /* Here we assume SIMD-enabled kernels are being used. But as currently
380 * calc_verlet_buffer_size gives the same results for 4x8 and 4x4
381 * and 4x2 gives a larger buffer than 4x4, this is ok.
383 ListSetupType listType = (makeGpuPairList ? ListSetupType::Gpu : ListSetupType::CpuSimdWhenSupported);
384 VerletbufListSetup listSetup = verletbufGetSafeListSetup(listType);
386 const real rlist_new =
387 calcVerletBufferSize(*mtop, det(box), *ir, ir->nstlist, ir->nstlist - 1, -1, listSetup);
389 if (rlist_new != ir->rlist)
391 if (fplog != nullptr)
393 fprintf(fplog, "\nChanging rlist from %g to %g for non-bonded %dx%d atom kernels\n\n",
394 ir->rlist, rlist_new,
395 listSetup.cluster_size_i, listSetup.cluster_size_j);
397 ir->rlist = rlist_new;
401 if (nstlist_cmdline > 0 && (!EI_DYNAMICS(ir->eI) || ir->verletbuf_tol <= 0))
403 gmx_fatal(FARGS, "Can not set nstlist without %s",
404 !EI_DYNAMICS(ir->eI) ? "dynamics" : "verlet-buffer-tolerance");
407 if (EI_DYNAMICS(ir->eI))
409 /* Set or try nstlist values */
410 increaseNstlist(fplog, cr, ir, nstlist_cmdline, mtop, box, makeGpuPairList, cpuinfo);
414 /*! \brief Override the nslist value in inputrec
416 * with value passed on the command line (if any)
418 static void override_nsteps_cmdline(const gmx::MDLogger &mdlog,
419 int64_t nsteps_cmdline,
424 /* override with anything else than the default -2 */
425 if (nsteps_cmdline > -2)
427 char sbuf_steps[STEPSTRSIZE];
428 char sbuf_msg[STRLEN];
430 ir->nsteps = nsteps_cmdline;
431 if (EI_DYNAMICS(ir->eI) && nsteps_cmdline != -1)
433 sprintf(sbuf_msg, "Overriding nsteps with value passed on the command line: %s steps, %.3g ps",
434 gmx_step_str(nsteps_cmdline, sbuf_steps),
435 fabs(nsteps_cmdline*ir->delta_t));
439 sprintf(sbuf_msg, "Overriding nsteps with value passed on the command line: %s steps",
440 gmx_step_str(nsteps_cmdline, sbuf_steps));
443 GMX_LOG(mdlog.warning).asParagraph().appendText(sbuf_msg);
445 else if (nsteps_cmdline < -2)
447 gmx_fatal(FARGS, "Invalid nsteps value passed on the command line: %" PRId64,
450 /* Do nothing if nsteps_cmdline == -2 */
456 /*! \brief Return whether GPU acceleration of nonbondeds is supported with the given settings.
458 * If not, and if a warning may be issued, logs a warning about
459 * falling back to CPU code. With thread-MPI, only the first
460 * call to this function should have \c issueWarning true. */
461 static bool gpuAccelerationOfNonbondedIsUseful(const MDLogger &mdlog,
462 const t_inputrec &ir,
465 bool gpuIsUseful = true;
468 if (ir.opts.ngener - ir.nwall > 1)
470 /* The GPU code does not support more than one energy group.
471 * If the user requested GPUs explicitly, a fatal error is given later.
475 "Multiple energy groups is not implemented for GPUs, falling back to the CPU. "
476 "For better performance, run on the GPU without energy groups and then do "
477 "gmx mdrun -rerun option on the trajectory with an energy group .tpr file.";
483 warning = "TPI is not implemented for GPUs.";
486 if (!gpuIsUseful && issueWarning)
488 GMX_LOG(mdlog.warning).asParagraph().appendText(warning);
494 //! Initializes the logger for mdrun.
495 static gmx::LoggerOwner buildLogger(FILE *fplog,
496 const bool isSimulationMasterRank)
498 gmx::LoggerBuilder builder;
499 if (fplog != nullptr)
501 builder.addTargetFile(gmx::MDLogger::LogLevel::Info, fplog);
503 if (isSimulationMasterRank)
505 builder.addTargetStream(gmx::MDLogger::LogLevel::Warning,
506 &gmx::TextOutputFile::standardError());
508 return builder.build();
511 //! Make a TaskTarget from an mdrun argument string.
512 static TaskTarget findTaskTarget(const char *optionString)
514 TaskTarget returnValue = TaskTarget::Auto;
516 if (strncmp(optionString, "auto", 3) == 0)
518 returnValue = TaskTarget::Auto;
520 else if (strncmp(optionString, "cpu", 3) == 0)
522 returnValue = TaskTarget::Cpu;
524 else if (strncmp(optionString, "gpu", 3) == 0)
526 returnValue = TaskTarget::Gpu;
530 GMX_ASSERT(false, "Option string should have been checked for sanity already");
536 //! Finish run, aggregate data to print performance info.
537 static void finish_run(FILE *fplog,
538 const gmx::MDLogger &mdlog,
540 const t_inputrec *inputrec,
541 t_nrnb nrnb[], gmx_wallcycle_t wcycle,
542 gmx_walltime_accounting_t walltime_accounting,
543 nonbonded_verlet_t *nbv,
544 const gmx_pme_t *pme,
548 double nbfs = 0, mflop = 0;
550 elapsed_time_over_all_ranks,
551 elapsed_time_over_all_threads,
552 elapsed_time_over_all_threads_over_all_ranks;
553 /* Control whether it is valid to print a report. Only the
554 simulation master may print, but it should not do so if the run
555 terminated e.g. before a scheduled reset step. This is
556 complicated by the fact that PME ranks are unaware of the
557 reason why they were sent a pmerecvqxFINISH. To avoid
558 communication deadlocks, we always do the communication for the
559 report, even if we've decided not to write the report, because
560 how long it takes to finish the run is not important when we've
561 decided not to report on the simulation performance.
563 Further, we only report performance for dynamical integrators,
564 because those are the only ones for which we plan to
565 consider doing any optimizations. */
566 bool printReport = EI_DYNAMICS(inputrec->eI) && SIMMASTER(cr);
568 if (printReport && !walltime_accounting_get_valid_finish(walltime_accounting))
570 GMX_LOG(mdlog.warning).asParagraph().appendText("Simulation ended prematurely, no performance report will be written.");
575 std::unique_ptr<t_nrnb> nrnbTotalStorage;
578 nrnbTotalStorage = std::make_unique<t_nrnb>();
579 nrnb_tot = nrnbTotalStorage.get();
581 MPI_Allreduce(nrnb->n, nrnb_tot->n, eNRNB, MPI_DOUBLE, MPI_SUM,
590 elapsed_time = walltime_accounting_get_time_since_reset(walltime_accounting);
591 elapsed_time_over_all_threads = walltime_accounting_get_time_since_reset_over_all_threads(walltime_accounting);
595 /* reduce elapsed_time over all MPI ranks in the current simulation */
596 MPI_Allreduce(&elapsed_time,
597 &elapsed_time_over_all_ranks,
598 1, MPI_DOUBLE, MPI_SUM,
600 elapsed_time_over_all_ranks /= cr->nnodes;
601 /* Reduce elapsed_time_over_all_threads over all MPI ranks in the
602 * current simulation. */
603 MPI_Allreduce(&elapsed_time_over_all_threads,
604 &elapsed_time_over_all_threads_over_all_ranks,
605 1, MPI_DOUBLE, MPI_SUM,
611 elapsed_time_over_all_ranks = elapsed_time;
612 elapsed_time_over_all_threads_over_all_ranks = elapsed_time_over_all_threads;
617 print_flop(fplog, nrnb_tot, &nbfs, &mflop);
620 if (thisRankHasDuty(cr, DUTY_PP) && DOMAINDECOMP(cr))
622 print_dd_statistics(cr, inputrec, fplog);
625 /* TODO Move the responsibility for any scaling by thread counts
626 * to the code that handled the thread region, so that there's a
627 * mechanism to keep cycle counting working during the transition
628 * to task parallelism. */
629 int nthreads_pp = gmx_omp_nthreads_get(emntNonbonded);
630 int nthreads_pme = gmx_omp_nthreads_get(emntPME);
631 wallcycle_scale_by_num_threads(wcycle, thisRankHasDuty(cr, DUTY_PME) && !thisRankHasDuty(cr, DUTY_PP), nthreads_pp, nthreads_pme);
632 auto cycle_sum(wallcycle_sum(cr, wcycle));
636 auto nbnxn_gpu_timings = (nbv != nullptr && nbv->useGpu()) ? Nbnxm::gpu_get_timings(nbv->gpu_nbv) : nullptr;
637 gmx_wallclock_gpu_pme_t pme_gpu_timings = {};
639 if (pme_gpu_task_enabled(pme))
641 pme_gpu_get_timings(pme, &pme_gpu_timings);
643 wallcycle_print(fplog, mdlog, cr->nnodes, cr->npmenodes, nthreads_pp, nthreads_pme,
644 elapsed_time_over_all_ranks,
649 if (EI_DYNAMICS(inputrec->eI))
651 delta_t = inputrec->delta_t;
656 print_perf(fplog, elapsed_time_over_all_threads_over_all_ranks,
657 elapsed_time_over_all_ranks,
658 walltime_accounting_get_nsteps_done_since_reset(walltime_accounting),
659 delta_t, nbfs, mflop);
663 print_perf(stderr, elapsed_time_over_all_threads_over_all_ranks,
664 elapsed_time_over_all_ranks,
665 walltime_accounting_get_nsteps_done_since_reset(walltime_accounting),
666 delta_t, nbfs, mflop);
671 int Mdrunner::mdrunner()
674 t_forcerec *fr = nullptr;
675 t_fcdata *fcd = nullptr;
676 real ewaldcoeff_q = 0;
677 real ewaldcoeff_lj = 0;
678 int nChargePerturbed = -1, nTypePerturbed = 0;
679 gmx_wallcycle_t wcycle;
680 gmx_walltime_accounting_t walltime_accounting = nullptr;
681 gmx_membed_t * membed = nullptr;
682 gmx_hw_info_t *hwinfo = nullptr;
684 /* CAUTION: threads may be started later on in this function, so
685 cr doesn't reflect the final parallel state right now */
688 /* TODO: inputrec should tell us whether we use an algorithm, not a file option */
689 const bool doEssentialDynamics = opt2bSet("-ei", filenames.size(), filenames.data());
690 const bool doMembed = opt2bSet("-membed", filenames.size(), filenames.data());
691 const bool doRerun = mdrunOptions.rerun;
693 // Handle task-assignment related user options.
694 EmulateGpuNonbonded emulateGpuNonbonded = (getenv("GMX_EMULATE_GPU") != nullptr ?
695 EmulateGpuNonbonded::Yes : EmulateGpuNonbonded::No);
697 std::vector<int> userGpuTaskAssignment;
700 userGpuTaskAssignment = parseUserTaskAssignmentString(hw_opt.userGpuTaskAssignment);
702 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
703 auto nonbondedTarget = findTaskTarget(nbpu_opt);
704 auto pmeTarget = findTaskTarget(pme_opt);
705 auto pmeFftTarget = findTaskTarget(pme_fft_opt);
706 auto bondedTarget = findTaskTarget(bonded_opt);
707 auto updateTarget = findTaskTarget(update_opt);
708 PmeRunMode pmeRunMode = PmeRunMode::None;
710 FILE *fplog = nullptr;
711 // If we are appending, we don't write log output because we need
712 // to check that the old log file matches what the checkpoint file
713 // expects. Otherwise, we should start to write log output now if
714 // there is a file ready for it.
715 if (logFileHandle != nullptr && startingBehavior != StartingBehavior::RestartWithAppending)
717 fplog = gmx_fio_getfp(logFileHandle);
719 const bool isSimulationMasterRank = findIsSimulationMasterRank(ms, communicator);
720 gmx::LoggerOwner logOwner(buildLogger(fplog, isSimulationMasterRank));
721 gmx::MDLogger mdlog(logOwner.logger());
723 // TODO The thread-MPI master rank makes a working
724 // PhysicalNodeCommunicator here, but it gets rebuilt by all ranks
725 // after the threads have been launched. This works because no use
726 // is made of that communicator until after the execution paths
727 // have rejoined. But it is likely that we can improve the way
728 // this is expressed, e.g. by expressly running detection only the
729 // master rank for thread-MPI, rather than relying on the mutex
730 // and reference count.
731 PhysicalNodeCommunicator physicalNodeComm(communicator, gmx_physicalnode_id_hash());
732 hwinfo = gmx_detect_hardware(mdlog, physicalNodeComm);
734 gmx_print_detected_hardware(fplog, isSimulationMasterRank && isMasterSim(ms), mdlog, hwinfo);
736 std::vector<int> gpuIdsToUse = makeGpuIdsToUse(hwinfo->gpu_info, hw_opt.gpuIdsAvailable);
738 // Print citation requests after all software/hardware printing
739 pleaseCiteGromacs(fplog);
741 // TODO Replace this by unique_ptr once t_inputrec is C++
742 t_inputrec inputrecInstance;
743 t_inputrec *inputrec = nullptr;
744 std::unique_ptr<t_state> globalState;
746 auto partialDeserializedTpr = std::make_unique<PartialDeserializedTprFile>();
748 if (isSimulationMasterRank)
750 /* Only the master rank has the global state */
751 globalState = std::make_unique<t_state>();
753 /* Read (nearly) all data required for the simulation
754 * and keep the partly serialized tpr contents to send to other ranks later
756 *partialDeserializedTpr = read_tpx_state(ftp2fn(efTPR, filenames.size(), filenames.data()), &inputrecInstance, globalState.get(), &mtop);
757 inputrec = &inputrecInstance;
760 /* Check and update the hardware options for internal consistency */
761 checkAndUpdateHardwareOptions(mdlog, &hw_opt, isSimulationMasterRank, domdecOptions.numPmeRanks, inputrec);
763 if (GMX_THREAD_MPI && isSimulationMasterRank)
765 bool useGpuForNonbonded = false;
766 bool useGpuForPme = false;
769 GMX_RELEASE_ASSERT(inputrec != nullptr, "Keep the compiler happy");
771 // If the user specified the number of ranks, then we must
772 // respect that, but in default mode, we need to allow for
773 // the number of GPUs to choose the number of ranks.
774 auto canUseGpuForNonbonded = buildSupportsNonbondedOnGpu(nullptr);
775 useGpuForNonbonded = decideWhetherToUseGpusForNonbondedWithThreadMpi
776 (nonbondedTarget, gpuIdsToUse, userGpuTaskAssignment, emulateGpuNonbonded,
777 canUseGpuForNonbonded,
778 gpuAccelerationOfNonbondedIsUseful(mdlog, *inputrec, GMX_THREAD_MPI),
779 hw_opt.nthreads_tmpi);
780 useGpuForPme = decideWhetherToUseGpusForPmeWithThreadMpi
781 (useGpuForNonbonded, pmeTarget, gpuIdsToUse, userGpuTaskAssignment,
782 *hwinfo, *inputrec, mtop, hw_opt.nthreads_tmpi, domdecOptions.numPmeRanks);
785 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
787 /* Determine how many thread-MPI ranks to start.
789 * TODO Over-writing the user-supplied value here does
790 * prevent any possible subsequent checks from working
792 hw_opt.nthreads_tmpi = get_nthreads_mpi(hwinfo,
801 // Now start the threads for thread MPI.
802 spawnThreads(hw_opt.nthreads_tmpi);
803 // The spawned threads enter mdrunner() and execution of
804 // master and spawned threads joins at the end of this block.
805 physicalNodeComm = PhysicalNodeCommunicator(communicator, gmx_physicalnode_id_hash());
808 GMX_RELEASE_ASSERT(communicator == MPI_COMM_WORLD, "Must have valid world communicator");
809 CommrecHandle crHandle = init_commrec(communicator, ms);
810 t_commrec *cr = crHandle.get();
811 GMX_RELEASE_ASSERT(cr != nullptr, "Must have valid commrec");
815 /* now broadcast everything to the non-master nodes/threads: */
816 if (!isSimulationMasterRank)
818 inputrec = &inputrecInstance;
820 init_parallel(cr, inputrec, &mtop, partialDeserializedTpr.get());
822 GMX_RELEASE_ASSERT(inputrec != nullptr, "All ranks should have a valid inputrec now");
823 partialDeserializedTpr.reset(nullptr);
825 // Now the number of ranks is known to all ranks, and each knows
826 // the inputrec read by the master rank. The ranks can now all run
827 // the task-deciding functions and will agree on the result
828 // without needing to communicate.
830 // TODO Should we do the communication in debug mode to support
831 // having an assertion?
833 // Note that these variables describe only their own node.
835 // Note that when bonded interactions run on a GPU they always run
836 // alongside a nonbonded task, so do not influence task assignment
837 // even though they affect the force calculation workload.
838 bool useGpuForNonbonded = false;
839 bool useGpuForPme = false;
840 bool useGpuForBonded = false;
841 bool useGpuForUpdate = false;
842 bool gpusWereDetected = hwinfo->ngpu_compatible_tot > 0;
845 // It's possible that there are different numbers of GPUs on
846 // different nodes, which is the user's responsibility to
847 // handle. If unsuitable, we will notice that during task
849 auto canUseGpuForNonbonded = buildSupportsNonbondedOnGpu(nullptr);
850 useGpuForNonbonded = decideWhetherToUseGpusForNonbonded(nonbondedTarget, userGpuTaskAssignment,
852 canUseGpuForNonbonded,
853 gpuAccelerationOfNonbondedIsUseful(mdlog, *inputrec, !GMX_THREAD_MPI),
855 useGpuForPme = decideWhetherToUseGpusForPme(useGpuForNonbonded, pmeTarget, userGpuTaskAssignment,
856 *hwinfo, *inputrec, mtop,
857 cr->nnodes, domdecOptions.numPmeRanks,
859 auto canUseGpuForBonded = buildSupportsGpuBondeds(nullptr) && inputSupportsGpuBondeds(*inputrec, mtop, nullptr);
861 decideWhetherToUseGpusForBonded(useGpuForNonbonded, useGpuForPme,
862 bondedTarget, canUseGpuForBonded,
863 EVDW_PME(inputrec->vdwtype),
864 EEL_PME_EWALD(inputrec->coulombtype),
865 domdecOptions.numPmeRanks, gpusWereDetected);
867 pmeRunMode = (useGpuForPme ? PmeRunMode::GPU : PmeRunMode::CPU);
868 if (pmeRunMode == PmeRunMode::GPU)
870 if (pmeFftTarget == TaskTarget::Cpu)
872 pmeRunMode = PmeRunMode::Mixed;
875 else if (pmeFftTarget == TaskTarget::Gpu)
877 gmx_fatal(FARGS, "Assigning FFTs to GPU requires PME to be assigned to GPU as well. With PME on CPU you should not be using -pmefft.");
880 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
882 // Initialize development feature flags that enabled by environment variable
883 // and report those features that are enabled.
884 const DevelopmentFeatureFlags devFlags = manageDevelopmentFeatures(mdlog, useGpuForNonbonded, useGpuForPme);
887 // TODO: hide restraint implementation details from Mdrunner.
888 // There is nothing unique about restraints at this point as far as the
889 // Mdrunner is concerned. The Mdrunner should just be getting a sequence of
890 // factory functions from the SimulationContext on which to call mdModules_->add().
891 // TODO: capture all restraints into a single RestraintModule, passed to the runner builder.
892 for (auto && restraint : restraintManager_->getRestraints())
894 auto module = RestraintMDModule::create(restraint,
896 mdModules_->add(std::move(module));
899 // TODO: Error handling
900 mdModules_->assignOptionsToModules(*inputrec->params, nullptr);
901 const auto &mdModulesNotifier = mdModules_->notifier().notifier_;
903 if (inputrec->internalParameters != nullptr)
905 mdModulesNotifier.notify(*inputrec->internalParameters);
908 if (fplog != nullptr)
910 pr_inputrec(fplog, 0, "Input Parameters", inputrec, FALSE);
911 fprintf(fplog, "\n");
916 /* In rerun, set velocities to zero if present */
917 if (doRerun && ((globalState->flags & (1 << estV)) != 0))
919 // rerun does not use velocities
920 GMX_LOG(mdlog.info).asParagraph().appendText(
921 "Rerun trajectory contains velocities. Rerun does only evaluate "
922 "potential energy and forces. The velocities will be ignored.");
923 for (int i = 0; i < globalState->natoms; i++)
925 clear_rvec(globalState->v[i]);
927 globalState->flags &= ~(1 << estV);
930 /* now make sure the state is initialized and propagated */
931 set_state_entries(globalState.get(), inputrec);
934 /* NM and TPI parallelize over force/energy calculations, not atoms,
935 * so we need to initialize and broadcast the global state.
937 if (inputrec->eI == eiNM || inputrec->eI == eiTPI)
941 globalState = std::make_unique<t_state>();
943 broadcastStateWithoutDynamics(cr, globalState.get());
946 /* A parallel command line option consistency check that we can
947 only do after any threads have started. */
948 if (!PAR(cr) && (domdecOptions.numCells[XX] > 1 ||
949 domdecOptions.numCells[YY] > 1 ||
950 domdecOptions.numCells[ZZ] > 1 ||
951 domdecOptions.numPmeRanks > 0))
954 "The -dd or -npme option request a parallel simulation, "
956 "but %s was compiled without threads or MPI enabled", output_env_get_program_display_name(oenv));
958 "but the number of MPI-threads (option -ntmpi) is not set or is 1");
960 "but %s was not started through mpirun/mpiexec or only one rank was requested through mpirun/mpiexec", output_env_get_program_display_name(oenv));
965 (EI_ENERGY_MINIMIZATION(inputrec->eI) || eiNM == inputrec->eI))
967 gmx_fatal(FARGS, "The .mdp file specified an energy mininization or normal mode algorithm, and these are not compatible with mdrun -rerun");
970 if (!(EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype)))
972 if (domdecOptions.numPmeRanks > 0)
974 gmx_fatal_collective(FARGS, cr->mpi_comm_mysim, MASTER(cr),
975 "PME-only ranks are requested, but the system does not use PME 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, "PME GPU decomposition is not supported");
1005 fcRegisterSteps(inputrec->nsteps, inputrec->init_step);
1009 /* NMR restraints must be initialized before load_checkpoint,
1010 * since with time averaging the history is added to t_state.
1011 * For proper consistency check we therefore need to extend
1013 * So the PME-only nodes (if present) will also initialize
1014 * the distance restraints.
1018 /* This needs to be called before read_checkpoint to extend the state */
1019 init_disres(fplog, &mtop, inputrec, cr, ms, fcd, globalState.get(), replExParams.exchangeInterval > 0);
1021 init_orires(fplog, &mtop, inputrec, cr, ms, globalState.get(), &(fcd->orires));
1023 auto deform = prepareBoxDeformation(globalState->box, cr, *inputrec);
1025 ObservablesHistory observablesHistory = {};
1027 if (startingBehavior != StartingBehavior::NewSimulation)
1029 /* Check if checkpoint file exists before doing continuation.
1030 * This way we can use identical input options for the first and subsequent runs...
1032 if (mdrunOptions.numStepsCommandline > -2)
1034 /* Temporarily set the number of steps to unlimited to avoid
1035 * triggering the nsteps check in load_checkpoint().
1036 * This hack will go away soon when the -nsteps option is removed.
1038 inputrec->nsteps = -1;
1041 load_checkpoint(opt2fn_master("-cpi", filenames.size(), filenames.data(), cr),
1043 cr, domdecOptions.numCells,
1044 inputrec, globalState.get(),
1045 &observablesHistory,
1046 mdrunOptions.reproducible, mdModules_->notifier());
1048 if (startingBehavior == StartingBehavior::RestartWithAppending && logFileHandle)
1050 // Now we can start normal logging to the truncated log file.
1051 fplog = gmx_fio_getfp(logFileHandle);
1052 prepareLogAppending(fplog);
1053 logOwner = buildLogger(fplog, MASTER(cr));
1054 mdlog = logOwner.logger();
1058 if (mdrunOptions.numStepsCommandline > -2)
1060 GMX_LOG(mdlog.info).asParagraph().
1061 appendText("The -nsteps functionality is deprecated, and may be removed in a future version. "
1062 "Consider using gmx convert-tpr -nsteps or changing the appropriate .mdp file field.");
1064 /* override nsteps with value set on the commandline */
1065 override_nsteps_cmdline(mdlog, mdrunOptions.numStepsCommandline, inputrec);
1069 copy_mat(globalState->box, box);
1074 gmx_bcast(sizeof(box), box, cr);
1077 if (inputrec->cutoff_scheme != ecutsVERLET)
1079 gmx_fatal(FARGS, "This group-scheme .tpr file can no longer be run by mdrun. Please update to the Verlet scheme, or use an earlier version of GROMACS if necessary.");
1081 /* Update rlist and nstlist. */
1082 prepare_verlet_scheme(fplog, cr, inputrec, nstlist_cmdline, &mtop, box,
1083 useGpuForNonbonded || (emulateGpuNonbonded == EmulateGpuNonbonded::Yes), *hwinfo->cpuInfo);
1085 const bool prefer1DAnd1PulseDD = (devFlags.enableGpuHaloExchange && useGpuForNonbonded);
1086 // This builder is necessary while we have multi-part construction
1087 // of DD. Before DD is constructed, we use the existence of
1088 // the builder object to indicate that further construction of DD
1090 std::unique_ptr<DomainDecompositionBuilder> ddBuilder;
1091 if (PAR(cr) && !(EI_TPI(inputrec->eI) ||
1092 inputrec->eI == eiNM))
1094 ddBuilder = std::make_unique<DomainDecompositionBuilder>
1095 (mdlog, cr, domdecOptions, mdrunOptions,
1096 prefer1DAnd1PulseDD, mtop, *inputrec,
1097 box, positionsFromStatePointer(globalState.get()));
1101 /* PME, if used, is done on all nodes with 1D decomposition */
1103 cr->duty = (DUTY_PP | DUTY_PME);
1105 if (inputrec->ePBC == epbcSCREW)
1108 "pbc=screw is only implemented with domain decomposition");
1112 // Produce the task assignment for this rank.
1113 GpuTaskAssignmentsBuilder gpuTaskAssignmentsBuilder;
1114 GpuTaskAssignments gpuTaskAssignments =
1115 gpuTaskAssignmentsBuilder.build(gpuIdsToUse,
1116 userGpuTaskAssignment,
1127 thisRankHasDuty(cr, DUTY_PP),
1128 // TODO cr->duty & DUTY_PME should imply that a PME
1129 // algorithm is active, but currently does not.
1130 EEL_PME(inputrec->coulombtype) &&
1131 thisRankHasDuty(cr, DUTY_PME));
1133 const bool printHostName = (cr->nnodes > 1);
1134 gpuTaskAssignments.reportGpuUsage(mdlog, printHostName, useGpuForBonded, pmeRunMode);
1136 // If the user chose a task assignment, give them some hints
1137 // where appropriate.
1138 if (!userGpuTaskAssignment.empty())
1140 gpuTaskAssignments.logPerformanceHints(mdlog,
1141 ssize(gpuIdsToUse));
1144 // Get the device handles for the modules, nullptr when no task is assigned.
1145 gmx_device_info_t *nonbondedDeviceInfo = gpuTaskAssignments.initNonbondedDevice(cr);
1146 gmx_device_info_t *pmeDeviceInfo = gpuTaskAssignments.initPmeDevice();
1148 // TODO Initialize GPU streams here.
1150 // TODO Currently this is always built, yet DD partition code
1151 // checks if it is built before using it. Probably it should
1152 // become an MDModule that is made only when another module
1153 // requires it (e.g. pull, CompEl, density fitting), so that we
1154 // don't update the local atom sets unilaterally every step.
1155 LocalAtomSetManager atomSets;
1158 // TODO Pass the GPU streams to ddBuilder to use in buffer
1159 // transfers (e.g. halo exchange)
1160 cr->dd = ddBuilder->build(&atomSets);
1161 // The builder's job is done, so destruct it
1162 ddBuilder.reset(nullptr);
1163 // Note that local state still does not exist yet.
1168 /* After possible communicator splitting in make_dd_communicators.
1169 * we can set up the intra/inter node communication.
1171 gmx_setup_nodecomm(fplog, cr);
1177 GMX_LOG(mdlog.warning).asParagraph().appendTextFormatted(
1178 "This is simulation %d out of %d running as a composite GROMACS\n"
1179 "multi-simulation job. Setup for this simulation:\n",
1182 GMX_LOG(mdlog.warning).appendTextFormatted(
1183 "Using %d MPI %s\n",
1186 cr->nnodes == 1 ? "thread" : "threads"
1188 cr->nnodes == 1 ? "process" : "processes"
1194 // If mdrun -pin auto honors any affinity setting that already
1195 // exists. If so, it is nice to provide feedback about whether
1196 // that existing affinity setting was from OpenMP or something
1197 // else, so we run this code both before and after we initialize
1198 // the OpenMP support.
1199 gmx_check_thread_affinity_set(mdlog,
1200 &hw_opt, hwinfo->nthreads_hw_avail, FALSE);
1201 /* Check and update the number of OpenMP threads requested */
1202 checkAndUpdateRequestedNumOpenmpThreads(&hw_opt, *hwinfo, cr, ms, physicalNodeComm.size_,
1203 pmeRunMode, mtop, *inputrec);
1205 gmx_omp_nthreads_init(mdlog, cr,
1206 hwinfo->nthreads_hw_avail,
1207 physicalNodeComm.size_,
1208 hw_opt.nthreads_omp,
1209 hw_opt.nthreads_omp_pme,
1210 !thisRankHasDuty(cr, DUTY_PP));
1212 // Enable FP exception detection, but not in
1213 // Release mode and not for compilers with known buggy FP
1214 // exception support (clang with any optimization) or suspected
1215 // buggy FP exception support (gcc 7.* with optimization).
1216 #if !defined NDEBUG && \
1217 !((defined __clang__ || (defined(__GNUC__) && !defined(__ICC) && __GNUC__ == 7)) \
1218 && defined __OPTIMIZE__)
1219 const bool bEnableFPE = true;
1221 const bool bEnableFPE = false;
1223 //FIXME - reconcile with gmx_feenableexcept() call from CommandLineModuleManager::run()
1226 gmx_feenableexcept();
1229 /* Now that we know the setup is consistent, check for efficiency */
1230 check_resource_division_efficiency(hwinfo,
1231 gpuTaskAssignments.thisRankHasAnyGpuTask(),
1232 mdrunOptions.ntompOptionIsSet,
1236 /* getting number of PP/PME threads on this MPI / tMPI rank.
1237 PME: env variable should be read only on one node to make sure it is
1238 identical everywhere;
1240 const int numThreadsOnThisRank =
1241 thisRankHasDuty(cr, DUTY_PP) ? gmx_omp_nthreads_get(emntNonbonded) : gmx_omp_nthreads_get(emntPME);
1242 checkHardwareOversubscription(numThreadsOnThisRank, cr->nodeid,
1243 *hwinfo->hardwareTopology,
1244 physicalNodeComm, mdlog);
1246 // Enable Peer access between GPUs where available
1247 // Only for DD, only master PP rank needs to perform setup, and only if thread MPI plus
1248 // any of the GPU communication features are active.
1249 if (DOMAINDECOMP(cr) && MASTER(cr) && thisRankHasDuty(cr, DUTY_PP) && GMX_THREAD_MPI &&
1250 (devFlags.enableGpuHaloExchange || devFlags.enableGpuPmePPComm))
1252 setupGpuDevicePeerAccess(gpuIdsToUse, mdlog);
1255 if (hw_opt.threadAffinity != ThreadAffinity::Off)
1257 /* Before setting affinity, check whether the affinity has changed
1258 * - which indicates that probably the OpenMP library has changed it
1259 * since we first checked).
1261 gmx_check_thread_affinity_set(mdlog,
1262 &hw_opt, hwinfo->nthreads_hw_avail, TRUE);
1264 int numThreadsOnThisNode, intraNodeThreadOffset;
1265 analyzeThreadsOnThisNode(physicalNodeComm, numThreadsOnThisRank, &numThreadsOnThisNode,
1266 &intraNodeThreadOffset);
1268 /* Set the CPU affinity */
1269 gmx_set_thread_affinity(mdlog, cr, &hw_opt, *hwinfo->hardwareTopology,
1270 numThreadsOnThisRank, numThreadsOnThisNode,
1271 intraNodeThreadOffset, nullptr);
1274 if (mdrunOptions.timingOptions.resetStep > -1)
1276 GMX_LOG(mdlog.info).asParagraph().
1277 appendText("The -resetstep functionality is deprecated, and may be removed in a future version.");
1279 wcycle = wallcycle_init(fplog, mdrunOptions.timingOptions.resetStep, cr);
1283 /* Master synchronizes its value of reset_counters with all nodes
1284 * including PME only nodes */
1285 int64_t reset_counters = wcycle_get_reset_counters(wcycle);
1286 gmx_bcast_sim(sizeof(reset_counters), &reset_counters, cr);
1287 wcycle_set_reset_counters(wcycle, reset_counters);
1290 // Membrane embedding must be initialized before we call init_forcerec()
1295 fprintf(stderr, "Initializing membed");
1297 /* Note that membed cannot work in parallel because mtop is
1298 * changed here. Fix this if we ever want to make it run with
1299 * multiple ranks. */
1300 membed = init_membed(fplog, filenames.size(), filenames.data(), &mtop, inputrec, globalState.get(), cr,
1302 .checkpointOptions.period);
1305 const bool thisRankHasPmeGpuTask = gpuTaskAssignments.thisRankHasPmeGpuTask();
1306 std::unique_ptr<MDAtoms> mdAtoms;
1307 std::unique_ptr<gmx_vsite_t> vsite;
1310 if (thisRankHasDuty(cr, DUTY_PP))
1312 mdModulesNotifier.notify(*cr);
1313 mdModulesNotifier.notify(&atomSets);
1314 mdModulesNotifier.notify(PeriodicBoundaryConditionType {inputrec->ePBC});
1315 mdModulesNotifier.notify(SimulationTimeStep { inputrec->delta_t });
1316 /* Initiate forcerecord */
1317 fr = new t_forcerec;
1318 fr->forceProviders = mdModules_->initForceProviders();
1319 init_forcerec(fplog, mdlog, fr, fcd,
1320 inputrec, &mtop, cr, box,
1321 opt2fn("-table", filenames.size(), filenames.data()),
1322 opt2fn("-tablep", filenames.size(), filenames.data()),
1323 opt2fns("-tableb", filenames.size(), filenames.data()),
1324 *hwinfo, nonbondedDeviceInfo,
1326 pmeRunMode == PmeRunMode::GPU && !thisRankHasDuty(cr, DUTY_PME),
1330 // TODO Move this to happen during domain decomposition setup,
1331 // once stream and event handling works well with that.
1332 // TODO remove need to pass local stream into GPU halo exchange - Redmine #3093
1333 if (havePPDomainDecomposition(cr) && prefer1DAnd1PulseDD && is1DAnd1PulseDD(*cr->dd))
1335 GMX_RELEASE_ASSERT(devFlags.enableGpuBufferOps, "Must use GMX_GPU_BUFFER_OPS=1 to use GMX_GPU_DD_COMMS=1");
1336 void *streamLocal = Nbnxm::gpu_get_command_stream(fr->nbv->gpu_nbv, Nbnxm::InteractionLocality::Local);
1337 void *streamNonLocal = Nbnxm::gpu_get_command_stream(fr->nbv->gpu_nbv, Nbnxm::InteractionLocality::NonLocal);
1338 void *coordinatesOnDeviceEvent = fr->nbv->get_x_on_device_event();
1339 GMX_LOG(mdlog.warning).asParagraph().appendTextFormatted(
1340 "NOTE: This run uses the 'GPU halo exchange' feature, enabled by the GMX_GPU_DD_COMMS environment variable.");
1341 cr->dd->gpuHaloExchange = std::make_unique<GpuHaloExchange>(cr->dd, cr->mpi_comm_mysim, streamLocal,
1342 streamNonLocal, coordinatesOnDeviceEvent);
1345 /* Initialize the mdAtoms structure.
1346 * mdAtoms is not filled with atom data,
1347 * as this can not be done now with domain decomposition.
1349 mdAtoms = makeMDAtoms(fplog, mtop, *inputrec, thisRankHasPmeGpuTask);
1350 if (globalState && thisRankHasPmeGpuTask)
1352 // The pinning of coordinates in the global state object works, because we only use
1353 // PME on GPU without DD or on a separate PME rank, and because the local state pointer
1354 // points to the global state object without DD.
1355 // FIXME: MD and EM separately set up the local state - this should happen in the same function,
1356 // which should also perform the pinning.
1357 changePinningPolicy(&globalState->x, pme_get_pinning_policy());
1360 /* Initialize the virtual site communication */
1361 vsite = initVsite(mtop, cr);
1363 calc_shifts(box, fr->shift_vec);
1365 /* With periodic molecules the charge groups should be whole at start up
1366 * and the virtual sites should not be far from their proper positions.
1368 if (!inputrec->bContinuation && MASTER(cr) &&
1369 !(inputrec->ePBC != epbcNONE && inputrec->bPeriodicMols))
1371 /* Make molecules whole at start of run */
1372 if (fr->ePBC != epbcNONE)
1374 do_pbc_first_mtop(fplog, inputrec->ePBC, box, &mtop, globalState->x.rvec_array());
1378 /* Correct initial vsite positions are required
1379 * for the initial distribution in the domain decomposition
1380 * and for the initial shell prediction.
1382 constructVsitesGlobal(mtop, globalState->x);
1386 if (EEL_PME(fr->ic->eeltype) || EVDW_PME(fr->ic->vdwtype))
1388 ewaldcoeff_q = fr->ic->ewaldcoeff_q;
1389 ewaldcoeff_lj = fr->ic->ewaldcoeff_lj;
1394 /* This is a PME only node */
1396 GMX_ASSERT(globalState == nullptr, "We don't need the state on a PME only rank and expect it to be unitialized");
1398 ewaldcoeff_q = calc_ewaldcoeff_q(inputrec->rcoulomb, inputrec->ewald_rtol);
1399 ewaldcoeff_lj = calc_ewaldcoeff_lj(inputrec->rvdw, inputrec->ewald_rtol_lj);
1402 gmx_pme_t *sepPmeData = nullptr;
1403 // This reference hides the fact that PME data is owned by runner on PME-only ranks and by forcerec on other ranks
1404 GMX_ASSERT(thisRankHasDuty(cr, DUTY_PP) == (fr != nullptr), "Double-checking that only PME-only ranks have no forcerec");
1405 gmx_pme_t * &pmedata = fr ? fr->pmedata : sepPmeData;
1407 // TODO should live in ewald module once its testing is improved
1409 // Later, this program could contain kernels that might be later
1410 // re-used as auto-tuning progresses, or subsequent simulations
1412 PmeGpuProgramStorage pmeGpuProgram;
1413 if (thisRankHasPmeGpuTask)
1415 pmeGpuProgram = buildPmeGpuProgram(pmeDeviceInfo);
1418 /* Initiate PME if necessary,
1419 * either on all nodes or on dedicated PME nodes only. */
1420 if (EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype))
1422 if (mdAtoms && mdAtoms->mdatoms())
1424 nChargePerturbed = mdAtoms->mdatoms()->nChargePerturbed;
1425 if (EVDW_PME(inputrec->vdwtype))
1427 nTypePerturbed = mdAtoms->mdatoms()->nTypePerturbed;
1430 if (cr->npmenodes > 0)
1432 /* The PME only nodes need to know nChargePerturbed(FEP on Q) and nTypePerturbed(FEP on LJ)*/
1433 gmx_bcast_sim(sizeof(nChargePerturbed), &nChargePerturbed, cr);
1434 gmx_bcast_sim(sizeof(nTypePerturbed), &nTypePerturbed, cr);
1437 if (thisRankHasDuty(cr, DUTY_PME))
1441 pmedata = gmx_pme_init(cr,
1442 getNumPmeDomains(cr->dd),
1444 nChargePerturbed != 0, nTypePerturbed != 0,
1445 mdrunOptions.reproducible,
1446 ewaldcoeff_q, ewaldcoeff_lj,
1447 gmx_omp_nthreads_get(emntPME),
1448 pmeRunMode, nullptr,
1449 pmeDeviceInfo, pmeGpuProgram.get(), mdlog);
1451 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1456 if (EI_DYNAMICS(inputrec->eI))
1458 /* Turn on signal handling on all nodes */
1460 * (A user signal from the PME nodes (if any)
1461 * is communicated to the PP nodes.
1463 signal_handler_install();
1466 pull_t *pull_work = nullptr;
1467 if (thisRankHasDuty(cr, DUTY_PP))
1469 /* Assumes uniform use of the number of OpenMP threads */
1470 walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(emntDefault));
1472 if (inputrec->bPull)
1474 /* Initialize pull code */
1476 init_pull(fplog, inputrec->pull, inputrec,
1477 &mtop, cr, &atomSets, inputrec->fepvals->init_lambda);
1478 if (inputrec->pull->bXOutAverage || inputrec->pull->bFOutAverage)
1480 initPullHistory(pull_work, &observablesHistory);
1482 if (EI_DYNAMICS(inputrec->eI) && MASTER(cr))
1484 init_pull_output_files(pull_work,
1485 filenames.size(), filenames.data(), oenv,
1490 std::unique_ptr<EnforcedRotation> enforcedRotation;
1493 /* Initialize enforced rotation code */
1494 enforcedRotation = init_rot(fplog,
1507 t_swap *swap = nullptr;
1508 if (inputrec->eSwapCoords != eswapNO)
1510 /* Initialize ion swapping code */
1511 swap = init_swapcoords(fplog, inputrec,
1512 opt2fn_master("-swap", filenames.size(), filenames.data(), cr),
1513 &mtop, globalState.get(), &observablesHistory,
1514 cr, &atomSets, oenv, mdrunOptions,
1518 /* Let makeConstraints know whether we have essential dynamics constraints. */
1519 auto constr = makeConstraints(mtop, *inputrec, pull_work, doEssentialDynamics,
1520 fplog, *mdAtoms->mdatoms(),
1521 cr, ms, &nrnb, wcycle, fr->bMolPBC);
1523 /* Energy terms and groups */
1524 gmx_enerdata_t enerd(mtop.groups.groups[SimulationAtomGroupType::EnergyOutput].size(), inputrec->fepvals->n_lambda);
1526 /* Kinetic energy data */
1527 gmx_ekindata_t ekind;
1528 init_ekindata(fplog, &mtop, &(inputrec->opts), &ekind, inputrec->cos_accel);
1530 /* Set up interactive MD (IMD) */
1531 auto imdSession = makeImdSession(inputrec, cr, wcycle, &enerd, ms, &mtop, mdlog,
1532 MASTER(cr) ? globalState->x.rvec_array() : nullptr,
1533 filenames.size(), filenames.data(), oenv, mdrunOptions.imdOptions,
1536 if (DOMAINDECOMP(cr))
1538 GMX_RELEASE_ASSERT(fr, "fr was NULL while cr->duty was DUTY_PP");
1539 /* This call is not included in init_domain_decomposition mainly
1540 * because fr->cginfo_mb is set later.
1542 dd_init_bondeds(fplog, cr->dd, &mtop, vsite.get(), inputrec,
1543 domdecOptions.checkBondedInteractions,
1547 if (updateTarget == TaskTarget::Gpu)
1551 gmx_fatal(FARGS, "It is currently not possible to redirect the calculation "
1552 "of update and constraints to the GPU!");
1556 // Before we start the actual simulator, try if we can run the update task on the GPU.
1557 useGpuForUpdate = decideWhetherToUseGpuForUpdate(devFlags.forceGpuUpdateDefaultOn,
1561 devFlags.enableGpuBufferOps,
1566 doEssentialDynamics,
1567 fcd->orires.nr != 0,
1568 fcd->disres.nsystems != 0,
1569 replExParams.exchangeInterval > 0);
1571 const bool inputIsCompatibleWithModularSimulator = ModularSimulator::isInputCompatible(
1573 inputrec, doRerun, vsite.get(), ms, replExParams,
1574 fcd, static_cast<int>(filenames.size()), filenames.data(),
1575 &observablesHistory, membed);
1577 const bool useModularSimulator = inputIsCompatibleWithModularSimulator && !(getenv("GMX_DISABLE_MODULAR_SIMULATOR") != nullptr);
1579 std::unique_ptr<gmx::StatePropagatorDataGpu> stateGpu;
1580 if (gpusWereDetected && ((useGpuForPme && thisRankHasDuty(cr, DUTY_PME)) || devFlags.enableGpuBufferOps))
1582 const void *pmeStream = pme_gpu_get_device_stream(fr->pmedata);
1583 const void *localStream = fr->nbv->gpu_nbv != nullptr ? Nbnxm::gpu_get_command_stream(fr->nbv->gpu_nbv, Nbnxm::InteractionLocality::Local) : nullptr;
1584 const void *nonLocalStream = fr->nbv->gpu_nbv != nullptr ? Nbnxm::gpu_get_command_stream(fr->nbv->gpu_nbv, Nbnxm::InteractionLocality::NonLocal) : nullptr;
1585 const void *deviceContext = pme_gpu_get_device_context(fr->pmedata);
1586 const int paddingSize = pme_gpu_get_padding_size(fr->pmedata);
1587 GpuApiCallBehavior transferKind = (inputrec->eI == eiMD && !doRerun && !useModularSimulator) ? GpuApiCallBehavior::Async : GpuApiCallBehavior::Sync;
1589 stateGpu = std::make_unique<gmx::StatePropagatorDataGpu>(pmeStream,
1595 fr->stateGpu = stateGpu.get();
1598 // TODO This is not the right place to manage the lifetime of
1599 // this data structure, but currently it's the easiest way to
1601 MdrunScheduleWorkload runScheduleWork;
1602 // Also populates the simulation constant workload description.
1603 runScheduleWork.simulationWork = createSimulationWorkload(useGpuForNonbonded,
1605 (pmeRunMode == PmeRunMode::GPU),
1608 devFlags.enableGpuBufferOps,
1609 devFlags.enableGpuHaloExchange,
1610 devFlags.enableGpuPmePPComm);
1613 GMX_ASSERT(stopHandlerBuilder_, "Runner must provide StopHandlerBuilder to simulator.");
1614 SimulatorBuilder simulatorBuilder;
1616 // build and run simulator object based on user-input
1617 auto simulator = simulatorBuilder.build(
1618 inputIsCompatibleWithModularSimulator,
1619 fplog, cr, ms, mdlog, static_cast<int>(filenames.size()), filenames.data(),
1623 vsite.get(), constr.get(),
1624 enforcedRotation ? enforcedRotation->getLegacyEnfrot() : nullptr,
1626 mdModules_->outputProvider(),
1627 mdModules_->notifier(),
1628 inputrec, imdSession.get(), pull_work, swap, &mtop,
1631 &observablesHistory,
1632 mdAtoms.get(), &nrnb, wcycle, fr,
1638 walltime_accounting,
1639 std::move(stopHandlerBuilder_),
1643 if (inputrec->bPull)
1645 finish_pull(pull_work);
1647 finish_swapcoords(swap);
1651 GMX_RELEASE_ASSERT(pmedata, "pmedata was NULL while cr->duty was not DUTY_PP");
1653 walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(emntPME));
1654 gmx_pmeonly(pmedata, cr, &nrnb, wcycle, walltime_accounting, inputrec, pmeRunMode);
1657 wallcycle_stop(wcycle, ewcRUN);
1659 /* Finish up, write some stuff
1660 * if rerunMD, don't write last frame again
1662 finish_run(fplog, mdlog, cr,
1663 inputrec, &nrnb, wcycle, walltime_accounting,
1664 fr ? fr->nbv.get() : nullptr,
1666 EI_DYNAMICS(inputrec->eI) && !isMultiSim(ms));
1668 // clean up cycle counter
1669 wallcycle_destroy(wcycle);
1674 gmx_pme_destroy(pmedata);
1678 // FIXME: this is only here to manually unpin mdAtoms->chargeA_ and state->x,
1679 // before we destroy the GPU context(s) in free_gpu_resources().
1680 // Pinned buffers are associated with contexts in CUDA.
1681 // As soon as we destroy GPU contexts after mdrunner() exits, these lines should go.
1682 mdAtoms.reset(nullptr);
1683 globalState.reset(nullptr);
1684 mdModules_.reset(nullptr); // destruct force providers here as they might also use the GPU
1686 /* Free GPU memory and set a physical node tMPI barrier (which should eventually go away) */
1687 free_gpu_resources(fr, physicalNodeComm, hwinfo->gpu_info);
1688 free_gpu(nonbondedDeviceInfo);
1689 free_gpu(pmeDeviceInfo);
1690 done_forcerec(fr, mtop.molblock.size());
1695 free_membed(membed);
1698 /* Does what it says */
1699 print_date_and_time(fplog, cr->nodeid, "Finished mdrun", gmx_gettime());
1700 walltime_accounting_destroy(walltime_accounting);
1702 // Ensure log file content is written
1705 gmx_fio_flush(logFileHandle);
1708 /* Reset FPEs (important for unit tests) by disabling them. Assumes no
1709 * exceptions were enabled before function was called. */
1712 gmx_fedisableexcept();
1715 auto rc = static_cast<int>(gmx_get_stop_condition());
1718 /* we need to join all threads. The sub-threads join when they
1719 exit this function, but the master thread needs to be told to
1721 if (PAR(cr) && MASTER(cr))
1729 Mdrunner::~Mdrunner()
1731 // Clean up of the Manager.
1732 // This will end up getting called on every thread-MPI rank, which is unnecessary,
1733 // but okay as long as threads synchronize some time before adding or accessing
1734 // a new set of restraints.
1735 if (restraintManager_)
1737 restraintManager_->clear();
1738 GMX_ASSERT(restraintManager_->countRestraints() == 0,
1739 "restraints added during runner life time should be cleared at runner destruction.");
1743 void Mdrunner::addPotential(std::shared_ptr<gmx::IRestraintPotential> puller,
1744 const std::string &name)
1746 GMX_ASSERT(restraintManager_, "Mdrunner must have a restraint manager.");
1747 // Not sure if this should be logged through the md logger or something else,
1748 // but it is helpful to have some sort of INFO level message sent somewhere.
1749 // std::cout << "Registering restraint named " << name << std::endl;
1751 // When multiple restraints are used, it may be wasteful to register them separately.
1752 // Maybe instead register an entire Restraint Manager as a force provider.
1753 restraintManager_->addToSpec(std::move(puller),
1757 Mdrunner::Mdrunner(std::unique_ptr<MDModules> mdModules)
1758 : mdModules_(std::move(mdModules))
1762 Mdrunner::Mdrunner(Mdrunner &&) noexcept = default;
1764 //NOLINTNEXTLINE(performance-noexcept-move-constructor) working around GCC bug 58265
1765 Mdrunner &Mdrunner::operator=(Mdrunner && /*handle*/) noexcept(BUGFREE_NOEXCEPT_STRING) = default;
1767 class Mdrunner::BuilderImplementation
1770 BuilderImplementation() = delete;
1771 BuilderImplementation(std::unique_ptr<MDModules> mdModules,
1772 compat::not_null<SimulationContext*> context);
1773 ~BuilderImplementation();
1775 BuilderImplementation &setExtraMdrunOptions(const MdrunOptions &options,
1776 real forceWarningThreshold,
1777 StartingBehavior startingBehavior);
1779 void addDomdec(const DomdecOptions &options);
1781 void addVerletList(int nstlist);
1783 void addReplicaExchange(const ReplicaExchangeParameters ¶ms);
1785 void addNonBonded(const char* nbpu_opt);
1787 void addPME(const char* pme_opt_, const char* pme_fft_opt_);
1789 void addBondedTaskAssignment(const char* bonded_opt);
1791 void addUpdateTaskAssignment(const char* update_opt);
1793 void addHardwareOptions(const gmx_hw_opt_t &hardwareOptions);
1795 void addFilenames(ArrayRef <const t_filenm> filenames);
1797 void addOutputEnvironment(gmx_output_env_t* outputEnvironment);
1799 void addLogFile(t_fileio *logFileHandle);
1801 void addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder);
1807 // Default parameters copied from runner.h
1808 // \todo Clarify source(s) of default parameters.
1810 const char* nbpu_opt_ = nullptr;
1811 const char* pme_opt_ = nullptr;
1812 const char* pme_fft_opt_ = nullptr;
1813 const char *bonded_opt_ = nullptr;
1814 const char *update_opt_ = nullptr;
1816 MdrunOptions mdrunOptions_;
1818 DomdecOptions domdecOptions_;
1820 ReplicaExchangeParameters replicaExchangeParameters_;
1822 //! Command-line override for the duration of a neighbor list with the Verlet scheme.
1825 //! Multisim communicator handle.
1826 gmx_multisim_t *multiSimulation_;
1828 //! mdrun communicator
1829 MPI_Comm communicator_ = MPI_COMM_NULL;
1831 //! Print a warning if any force is larger than this (in kJ/mol nm).
1832 real forceWarningThreshold_ = -1;
1834 //! Whether the simulation will start afresh, or restart with/without appending.
1835 StartingBehavior startingBehavior_ = StartingBehavior::NewSimulation;
1837 //! The modules that comprise the functionality of mdrun.
1838 std::unique_ptr<MDModules> mdModules_;
1840 //! \brief Parallelism information.
1841 gmx_hw_opt_t hardwareOptions_;
1843 //! filename options for simulation.
1844 ArrayRef<const t_filenm> filenames_;
1846 /*! \brief Handle to output environment.
1848 * \todo gmx_output_env_t needs lifetime management.
1850 gmx_output_env_t* outputEnvironment_ = nullptr;
1852 /*! \brief Non-owning handle to MD log file.
1854 * \todo Context should own output facilities for client.
1855 * \todo Improve log file handle management.
1857 * Code managing the FILE* relies on the ability to set it to
1858 * nullptr to check whether the filehandle is valid.
1860 t_fileio* logFileHandle_ = nullptr;
1863 * \brief Builder for simulation stop signal handler.
1865 std::unique_ptr<StopHandlerBuilder> stopHandlerBuilder_ = nullptr;
1868 Mdrunner::BuilderImplementation::BuilderImplementation(std::unique_ptr<MDModules> mdModules,
1869 compat::not_null<SimulationContext*> context) :
1870 mdModules_(std::move(mdModules))
1872 communicator_ = context->communicator_;
1873 multiSimulation_ = context->multiSimulation_.get();
1876 Mdrunner::BuilderImplementation::~BuilderImplementation() = default;
1878 Mdrunner::BuilderImplementation &
1879 Mdrunner::BuilderImplementation::setExtraMdrunOptions(const MdrunOptions &options,
1880 const real forceWarningThreshold,
1881 const StartingBehavior startingBehavior)
1883 mdrunOptions_ = options;
1884 forceWarningThreshold_ = forceWarningThreshold;
1885 startingBehavior_ = startingBehavior;
1889 void Mdrunner::BuilderImplementation::addDomdec(const DomdecOptions &options)
1891 domdecOptions_ = options;
1894 void Mdrunner::BuilderImplementation::addVerletList(int nstlist)
1899 void Mdrunner::BuilderImplementation::addReplicaExchange(const ReplicaExchangeParameters ¶ms)
1901 replicaExchangeParameters_ = params;
1904 Mdrunner Mdrunner::BuilderImplementation::build()
1906 auto newRunner = Mdrunner(std::move(mdModules_));
1908 newRunner.mdrunOptions = mdrunOptions_;
1909 newRunner.pforce = forceWarningThreshold_;
1910 newRunner.startingBehavior = startingBehavior_;
1911 newRunner.domdecOptions = domdecOptions_;
1913 // \todo determine an invariant to check or confirm that all gmx_hw_opt_t objects are valid
1914 newRunner.hw_opt = hardwareOptions_;
1916 // No invariant to check. This parameter exists to optionally override other behavior.
1917 newRunner.nstlist_cmdline = nstlist_;
1919 newRunner.replExParams = replicaExchangeParameters_;
1921 newRunner.filenames = filenames_;
1923 newRunner.communicator = communicator_;
1925 // nullptr is a valid value for the multisim handle
1926 newRunner.ms = multiSimulation_;
1928 // \todo Clarify ownership and lifetime management for gmx_output_env_t
1929 // \todo Update sanity checking when output environment has clearly specified invariants.
1930 // Initialization and default values for oenv are not well specified in the current version.
1931 if (outputEnvironment_)
1933 newRunner.oenv = outputEnvironment_;
1937 GMX_THROW(gmx::APIError("MdrunnerBuilder::addOutputEnvironment() is required before build()"));
1940 newRunner.logFileHandle = logFileHandle_;
1944 newRunner.nbpu_opt = nbpu_opt_;
1948 GMX_THROW(gmx::APIError("MdrunnerBuilder::addNonBonded() is required before build()"));
1951 if (pme_opt_ && pme_fft_opt_)
1953 newRunner.pme_opt = pme_opt_;
1954 newRunner.pme_fft_opt = pme_fft_opt_;
1958 GMX_THROW(gmx::APIError("MdrunnerBuilder::addElectrostatics() is required before build()"));
1963 newRunner.bonded_opt = bonded_opt_;
1967 GMX_THROW(gmx::APIError("MdrunnerBuilder::addBondedTaskAssignment() is required before build()"));
1972 newRunner.update_opt = update_opt_;
1976 GMX_THROW(gmx::APIError("MdrunnerBuilder::addUpdateTaskAssignment() is required before build() "));
1980 newRunner.restraintManager_ = std::make_unique<gmx::RestraintManager>();
1982 if (stopHandlerBuilder_)
1984 newRunner.stopHandlerBuilder_ = std::move(stopHandlerBuilder_);
1988 newRunner.stopHandlerBuilder_ = std::make_unique<StopHandlerBuilder>();
1994 void Mdrunner::BuilderImplementation::addNonBonded(const char* nbpu_opt)
1996 nbpu_opt_ = nbpu_opt;
1999 void Mdrunner::BuilderImplementation::addPME(const char* pme_opt,
2000 const char* pme_fft_opt)
2003 pme_fft_opt_ = pme_fft_opt;
2006 void Mdrunner::BuilderImplementation::addBondedTaskAssignment(const char* bonded_opt)
2008 bonded_opt_ = bonded_opt;
2011 void Mdrunner::BuilderImplementation::addUpdateTaskAssignment(const char* update_opt)
2013 update_opt_ = update_opt;
2016 void Mdrunner::BuilderImplementation::addHardwareOptions(const gmx_hw_opt_t &hardwareOptions)
2018 hardwareOptions_ = hardwareOptions;
2021 void Mdrunner::BuilderImplementation::addFilenames(ArrayRef<const t_filenm> filenames)
2023 filenames_ = filenames;
2026 void Mdrunner::BuilderImplementation::addOutputEnvironment(gmx_output_env_t* outputEnvironment)
2028 outputEnvironment_ = outputEnvironment;
2031 void Mdrunner::BuilderImplementation::addLogFile(t_fileio *logFileHandle)
2033 logFileHandle_ = logFileHandle;
2036 void Mdrunner::BuilderImplementation::addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder)
2038 stopHandlerBuilder_ = std::move(builder);
2041 MdrunnerBuilder::MdrunnerBuilder(std::unique_ptr<MDModules> mdModules,
2042 compat::not_null<SimulationContext*> context) :
2043 impl_ {std::make_unique<Mdrunner::BuilderImplementation>(std::move(mdModules), context)}
2047 MdrunnerBuilder::~MdrunnerBuilder() = default;
2049 MdrunnerBuilder &MdrunnerBuilder::addSimulationMethod(const MdrunOptions &options,
2050 real forceWarningThreshold,
2051 const StartingBehavior startingBehavior)
2053 impl_->setExtraMdrunOptions(options, forceWarningThreshold, startingBehavior);
2057 MdrunnerBuilder &MdrunnerBuilder::addDomainDecomposition(const DomdecOptions &options)
2059 impl_->addDomdec(options);
2063 MdrunnerBuilder &MdrunnerBuilder::addNeighborList(int nstlist)
2065 impl_->addVerletList(nstlist);
2069 MdrunnerBuilder &MdrunnerBuilder::addReplicaExchange(const ReplicaExchangeParameters ¶ms)
2071 impl_->addReplicaExchange(params);
2075 MdrunnerBuilder &MdrunnerBuilder::addNonBonded(const char* nbpu_opt)
2077 impl_->addNonBonded(nbpu_opt);
2081 MdrunnerBuilder &MdrunnerBuilder::addElectrostatics(const char* pme_opt,
2082 const char* pme_fft_opt)
2084 // The builder method may become more general in the future, but in this version,
2085 // parameters for PME electrostatics are both required and the only parameters
2087 if (pme_opt && pme_fft_opt)
2089 impl_->addPME(pme_opt, pme_fft_opt);
2093 GMX_THROW(gmx::InvalidInputError("addElectrostatics() arguments must be non-null pointers."));
2098 MdrunnerBuilder &MdrunnerBuilder::addBondedTaskAssignment(const char* bonded_opt)
2100 impl_->addBondedTaskAssignment(bonded_opt);
2104 MdrunnerBuilder &MdrunnerBuilder::addUpdateTaskAssignment(const char* update_opt)
2106 impl_->addUpdateTaskAssignment(update_opt);
2110 Mdrunner MdrunnerBuilder::build()
2112 return impl_->build();
2115 MdrunnerBuilder &MdrunnerBuilder::addHardwareOptions(const gmx_hw_opt_t &hardwareOptions)
2117 impl_->addHardwareOptions(hardwareOptions);
2121 MdrunnerBuilder &MdrunnerBuilder::addFilenames(ArrayRef<const t_filenm> filenames)
2123 impl_->addFilenames(filenames);
2127 MdrunnerBuilder &MdrunnerBuilder::addOutputEnvironment(gmx_output_env_t* outputEnvironment)
2129 impl_->addOutputEnvironment(outputEnvironment);
2133 MdrunnerBuilder &MdrunnerBuilder::addLogFile(t_fileio *logFileHandle)
2135 impl_->addLogFile(logFileHandle);
2139 MdrunnerBuilder &MdrunnerBuilder::addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder)
2141 impl_->addStopHandlerBuilder(std::move(builder));
2145 MdrunnerBuilder::MdrunnerBuilder(MdrunnerBuilder &&) noexcept = default;
2147 MdrunnerBuilder &MdrunnerBuilder::operator=(MdrunnerBuilder &&) noexcept = default;