2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
5 * Copyright (c) 2001-2004, The GROMACS development team.
6 * Copyright (c) 2011-2019,2020,2021, by the GROMACS development team, led by
7 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
8 * and including many others, as listed in the AUTHORS file in the
9 * top-level source directory and at http://www.gromacs.org.
11 * GROMACS is free software; you can redistribute it and/or
12 * modify it under the terms of the GNU Lesser General Public License
13 * as published by the Free Software Foundation; either version 2.1
14 * of the License, or (at your option) any later version.
16 * GROMACS is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
19 * Lesser General Public License for more details.
21 * You should have received a copy of the GNU Lesser General Public
22 * License along with GROMACS; if not, see
23 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
24 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
26 * If you want to redistribute modifications to GROMACS, please
27 * consider that scientific software is very special. Version
28 * control is crucial - bugs must be traceable. We will be happy to
29 * consider code for inclusion in the official distribution, but
30 * derived work must not be called official GROMACS. Details are found
31 * in the README & COPYING files - if they are missing, get the
32 * official version at http://www.gromacs.org.
34 * To help us fund GROMACS development, we humbly ask that you cite
35 * the research papers on the package. Check out http://www.gromacs.org.
39 * \brief Implements the MD runner routine calling all integrators.
41 * \author David van der Spoel <david.vanderspoel@icm.uu.se>
42 * \ingroup module_mdrun
59 #include "gromacs/commandline/filenm.h"
60 #include "gromacs/domdec/builder.h"
61 #include "gromacs/domdec/domdec.h"
62 #include "gromacs/domdec/domdec_struct.h"
63 #include "gromacs/domdec/gpuhaloexchange.h"
64 #include "gromacs/domdec/localatomsetmanager.h"
65 #include "gromacs/domdec/partition.h"
66 #include "gromacs/ewald/ewald_utils.h"
67 #include "gromacs/ewald/pme.h"
68 #include "gromacs/ewald/pme_gpu_program.h"
69 #include "gromacs/ewald/pme_only.h"
70 #include "gromacs/ewald/pme_pp_comm_gpu.h"
71 #include "gromacs/fileio/checkpoint.h"
72 #include "gromacs/fileio/gmxfio.h"
73 #include "gromacs/fileio/oenv.h"
74 #include "gromacs/fileio/tpxio.h"
75 #include "gromacs/gmxlib/network.h"
76 #include "gromacs/gmxlib/nrnb.h"
77 #include "gromacs/gpu_utils/device_stream_manager.h"
78 #include "gromacs/hardware/cpuinfo.h"
79 #include "gromacs/hardware/detecthardware.h"
80 #include "gromacs/hardware/device_management.h"
81 #include "gromacs/hardware/hardwaretopology.h"
82 #include "gromacs/hardware/printhardware.h"
83 #include "gromacs/imd/imd.h"
84 #include "gromacs/listed_forces/disre.h"
85 #include "gromacs/listed_forces/listed_forces_gpu.h"
86 #include "gromacs/listed_forces/listed_forces.h"
87 #include "gromacs/listed_forces/orires.h"
88 #include "gromacs/math/functions.h"
89 #include "gromacs/math/utilities.h"
90 #include "gromacs/math/vec.h"
91 #include "gromacs/mdlib/boxdeformation.h"
92 #include "gromacs/mdlib/broadcaststructs.h"
93 #include "gromacs/mdlib/calc_verletbuf.h"
94 #include "gromacs/mdlib/dispersioncorrection.h"
95 #include "gromacs/mdlib/enerdata_utils.h"
96 #include "gromacs/mdlib/force.h"
97 #include "gromacs/mdlib/forcerec.h"
98 #include "gromacs/mdlib/gmx_omp_nthreads.h"
99 #include "gromacs/mdlib/gpuforcereduction.h"
100 #include "gromacs/mdlib/makeconstraints.h"
101 #include "gromacs/mdlib/md_support.h"
102 #include "gromacs/mdlib/mdatoms.h"
103 #include "gromacs/mdlib/sighandler.h"
104 #include "gromacs/mdlib/stophandler.h"
105 #include "gromacs/mdlib/tgroup.h"
106 #include "gromacs/mdlib/updategroups.h"
107 #include "gromacs/mdlib/vsite.h"
108 #include "gromacs/mdrun/mdmodules.h"
109 #include "gromacs/mdrun/simulationcontext.h"
110 #include "gromacs/mdrun/simulationinput.h"
111 #include "gromacs/mdrun/simulationinputhandle.h"
112 #include "gromacs/mdrunutility/handlerestart.h"
113 #include "gromacs/mdrunutility/logging.h"
114 #include "gromacs/mdrunutility/multisim.h"
115 #include "gromacs/mdrunutility/printtime.h"
116 #include "gromacs/mdrunutility/threadaffinity.h"
117 #include "gromacs/mdtypes/checkpointdata.h"
118 #include "gromacs/mdtypes/commrec.h"
119 #include "gromacs/mdtypes/enerdata.h"
120 #include "gromacs/mdtypes/fcdata.h"
121 #include "gromacs/mdtypes/forcerec.h"
122 #include "gromacs/mdtypes/group.h"
123 #include "gromacs/mdtypes/inputrec.h"
124 #include "gromacs/mdtypes/interaction_const.h"
125 #include "gromacs/mdtypes/md_enums.h"
126 #include "gromacs/mdtypes/mdatom.h"
127 #include "gromacs/mdtypes/mdrunoptions.h"
128 #include "gromacs/mdtypes/observableshistory.h"
129 #include "gromacs/mdtypes/simulation_workload.h"
130 #include "gromacs/mdtypes/state.h"
131 #include "gromacs/mdtypes/state_propagator_data_gpu.h"
132 #include "gromacs/modularsimulator/modularsimulator.h"
133 #include "gromacs/nbnxm/gpu_data_mgmt.h"
134 #include "gromacs/nbnxm/nbnxm.h"
135 #include "gromacs/nbnxm/pairlist_tuning.h"
136 #include "gromacs/pbcutil/pbc.h"
137 #include "gromacs/pulling/output.h"
138 #include "gromacs/pulling/pull.h"
139 #include "gromacs/pulling/pull_rotation.h"
140 #include "gromacs/restraint/manager.h"
141 #include "gromacs/restraint/restraintmdmodule.h"
142 #include "gromacs/restraint/restraintpotential.h"
143 #include "gromacs/swap/swapcoords.h"
144 #include "gromacs/taskassignment/decidegpuusage.h"
145 #include "gromacs/taskassignment/decidesimulationworkload.h"
146 #include "gromacs/taskassignment/resourcedivision.h"
147 #include "gromacs/taskassignment/taskassignment.h"
148 #include "gromacs/taskassignment/usergpuids.h"
149 #include "gromacs/timing/gpu_timing.h"
150 #include "gromacs/timing/wallcycle.h"
151 #include "gromacs/timing/wallcyclereporting.h"
152 #include "gromacs/topology/mtop_util.h"
153 #include "gromacs/trajectory/trajectoryframe.h"
154 #include "gromacs/utility/basenetwork.h"
155 #include "gromacs/utility/cstringutil.h"
156 #include "gromacs/utility/exceptions.h"
157 #include "gromacs/utility/fatalerror.h"
158 #include "gromacs/utility/filestream.h"
159 #include "gromacs/utility/gmxassert.h"
160 #include "gromacs/utility/gmxmpi.h"
161 #include "gromacs/utility/keyvaluetree.h"
162 #include "gromacs/utility/logger.h"
163 #include "gromacs/utility/loggerbuilder.h"
164 #include "gromacs/utility/mdmodulesnotifiers.h"
165 #include "gromacs/utility/physicalnodecommunicator.h"
166 #include "gromacs/utility/pleasecite.h"
167 #include "gromacs/utility/programcontext.h"
168 #include "gromacs/utility/smalloc.h"
169 #include "gromacs/utility/stringutil.h"
170 #include "gromacs/utility/mpiinfo.h"
172 #include "isimulator.h"
173 #include "membedholder.h"
174 #include "replicaexchange.h"
175 #include "simulatorbuilder.h"
181 /*! \brief Manage any development feature flag variables encountered
183 * The use of dev features indicated by environment variables is
184 * logged in order to ensure that runs with such features enabled can
185 * be identified from their log and standard output. Any cross
186 * dependencies are also checked, and if unsatisfied, a fatal error
189 * Note that some development features overrides are applied already here:
190 * the GPU communication flags are set to false in non-tMPI and non-CUDA builds.
192 * \param[in] mdlog Logger object.
193 * \param[in] useGpuForNonbonded True if the nonbonded task is offloaded in this run.
194 * \param[in] pmeRunMode The PME run mode for this run
195 * \returns The object populated with development feature flags.
197 static DevelopmentFeatureFlags manageDevelopmentFeatures(const gmx::MDLogger& mdlog,
198 const bool useGpuForNonbonded,
199 const PmeRunMode pmeRunMode)
201 DevelopmentFeatureFlags devFlags;
203 // Some builds of GCC 5 give false positive warnings that these
204 // getenv results are ignored when clearly they are used.
205 #pragma GCC diagnostic push
206 #pragma GCC diagnostic ignored "-Wunused-result"
208 devFlags.enableGpuBufferOps =
209 GMX_GPU_CUDA && useGpuForNonbonded && (getenv("GMX_USE_GPU_BUFFER_OPS") != nullptr);
210 devFlags.enableGpuHaloExchange = GMX_GPU_CUDA && getenv("GMX_GPU_DD_COMMS") != nullptr;
211 devFlags.forceGpuUpdateDefault = (getenv("GMX_FORCE_UPDATE_DEFAULT_GPU") != nullptr) || GMX_FAHCORE;
212 devFlags.enableGpuPmePPComm = GMX_GPU_CUDA && getenv("GMX_GPU_PME_PP_COMMS") != nullptr;
214 #pragma GCC diagnostic pop
216 // Direct GPU comm path is being used with CUDA_AWARE_MPI
217 // make sure underlying MPI implementation is CUDA-aware
218 if (!GMX_THREAD_MPI && (devFlags.enableGpuPmePPComm || devFlags.enableGpuHaloExchange))
220 const bool haveDetectedCudaAwareMpi =
221 (checkMpiCudaAwareSupport() == CudaAwareMpiStatus::Supported);
222 const bool forceCudaAwareMpi = (getenv("GMX_FORCE_CUDA_AWARE_MPI") != nullptr);
224 if (!haveDetectedCudaAwareMpi && forceCudaAwareMpi)
226 // CUDA-aware support not detected in MPI library but, user has forced it's use
227 GMX_LOG(mdlog.warning)
229 .appendTextFormatted(
230 "This run has forced use of 'CUDA-aware MPI'. "
231 "But, GROMACS cannot determine if underlying MPI "
232 "is CUDA-aware. GROMACS recommends use of latest openMPI version "
233 "for CUDA-aware support. "
234 "If you observe failures at runtime, try unsetting "
235 "GMX_FORCE_CUDA_AWARE_MPI environment variable.");
238 if (haveDetectedCudaAwareMpi || forceCudaAwareMpi)
240 devFlags.usingCudaAwareMpi = true;
241 GMX_LOG(mdlog.warning)
243 .appendTextFormatted(
244 "Using CUDA-aware MPI for 'GPU halo exchange' or 'GPU PME-PP "
245 "communications' feature.");
249 if (devFlags.enableGpuHaloExchange)
251 GMX_LOG(mdlog.warning)
253 .appendTextFormatted(
254 "GMX_GPU_DD_COMMS environment variable detected, but the 'GPU "
255 "halo exchange' feature will not be enabled as GROMACS couldn't "
256 "detect CUDA_aware support in underlying MPI implementation.");
257 devFlags.enableGpuHaloExchange = false;
259 if (devFlags.enableGpuPmePPComm)
261 GMX_LOG(mdlog.warning)
264 "GMX_GPU_PME_PP_COMMS environment variable detected, but the "
265 "'GPU PME-PP communications' feature will not be enabled as "
267 "detect CUDA_aware support in underlying MPI implementation.");
268 devFlags.enableGpuPmePPComm = false;
271 GMX_LOG(mdlog.warning)
273 .appendTextFormatted(
274 "GROMACS recommends use of latest OpenMPI version for CUDA-aware "
276 "If you are certain about CUDA-aware support in your MPI library, "
277 "you can force it's use by setting environment variable "
278 " GMX_FORCE_CUDA_AWARE_MPI.");
282 if (devFlags.enableGpuBufferOps)
284 GMX_LOG(mdlog.warning)
286 .appendTextFormatted(
287 "This run uses the 'GPU buffer ops' feature, enabled by the "
288 "GMX_USE_GPU_BUFFER_OPS environment variable.");
291 if (devFlags.forceGpuUpdateDefault)
293 GMX_LOG(mdlog.warning)
295 .appendTextFormatted(
296 "This run will default to '-update gpu' as requested by the "
297 "GMX_FORCE_UPDATE_DEFAULT_GPU environment variable. GPU update with domain "
298 "decomposition lacks substantial testing and should be used with caution.");
301 if (devFlags.enableGpuHaloExchange)
303 if (useGpuForNonbonded)
305 if (!devFlags.enableGpuBufferOps)
307 GMX_LOG(mdlog.warning)
309 .appendTextFormatted(
310 "Enabling GPU buffer operations required by GMX_GPU_DD_COMMS "
311 "(equivalent with GMX_USE_GPU_BUFFER_OPS=1).");
312 devFlags.enableGpuBufferOps = true;
314 GMX_LOG(mdlog.warning)
316 .appendTextFormatted(
317 "This run has requested the 'GPU halo exchange' feature, enabled by "
319 "GMX_GPU_DD_COMMS environment variable.");
323 GMX_LOG(mdlog.warning)
325 .appendTextFormatted(
326 "GMX_GPU_DD_COMMS environment variable detected, but the 'GPU "
327 "halo exchange' feature will not be enabled as nonbonded interactions "
328 "are not offloaded.");
329 devFlags.enableGpuHaloExchange = false;
333 if (devFlags.enableGpuPmePPComm)
335 if (pmeRunMode == PmeRunMode::GPU)
337 if (!devFlags.enableGpuBufferOps)
339 GMX_LOG(mdlog.warning)
341 .appendTextFormatted(
342 "Enabling GPU buffer operations required by GMX_GPU_PME_PP_COMMS "
343 "(equivalent with GMX_USE_GPU_BUFFER_OPS=1).");
344 devFlags.enableGpuBufferOps = true;
346 GMX_LOG(mdlog.warning)
348 .appendTextFormatted(
349 "This run uses the 'GPU PME-PP communications' feature, enabled "
350 "by the GMX_GPU_PME_PP_COMMS environment variable.");
354 std::string clarification;
355 if (pmeRunMode == PmeRunMode::Mixed)
358 "PME FFT and gather are not offloaded to the GPU (PME is running in mixed "
363 clarification = "PME is not offloaded to the GPU.";
365 GMX_LOG(mdlog.warning)
368 "GMX_GPU_PME_PP_COMMS environment variable detected, but the "
369 "'GPU PME-PP communications' feature was not enabled as "
371 devFlags.enableGpuPmePPComm = false;
378 /*! \brief Barrier for safe simultaneous thread access to mdrunner data
380 * Used to ensure that the master thread does not modify mdrunner during copy
381 * on the spawned threads. */
382 static void threadMpiMdrunnerAccessBarrier()
385 MPI_Barrier(MPI_COMM_WORLD);
389 Mdrunner Mdrunner::cloneOnSpawnedThread() const
391 auto newRunner = Mdrunner(std::make_unique<MDModules>());
393 // All runners in the same process share a restraint manager resource because it is
394 // part of the interface to the client code, which is associated only with the
395 // original thread. Handles to the same resources can be obtained by copy.
397 newRunner.restraintManager_ = std::make_unique<RestraintManager>(*restraintManager_);
400 // Copy members of master runner.
401 // \todo Replace with builder when Simulation context and/or runner phases are better defined.
402 // Ref https://gitlab.com/gromacs/gromacs/-/issues/2587 and https://gitlab.com/gromacs/gromacs/-/issues/2375
403 newRunner.hw_opt = hw_opt;
404 newRunner.filenames = filenames;
406 newRunner.hwinfo_ = hwinfo_;
407 newRunner.oenv = oenv;
408 newRunner.mdrunOptions = mdrunOptions;
409 newRunner.domdecOptions = domdecOptions;
410 newRunner.nbpu_opt = nbpu_opt;
411 newRunner.pme_opt = pme_opt;
412 newRunner.pme_fft_opt = pme_fft_opt;
413 newRunner.bonded_opt = bonded_opt;
414 newRunner.update_opt = update_opt;
415 newRunner.nstlist_cmdline = nstlist_cmdline;
416 newRunner.replExParams = replExParams;
417 newRunner.pforce = pforce;
418 // Give the spawned thread the newly created valid communicator
419 // for the simulation.
420 newRunner.libraryWorldCommunicator = MPI_COMM_WORLD;
421 newRunner.simulationCommunicator = MPI_COMM_WORLD;
423 newRunner.startingBehavior = startingBehavior;
424 newRunner.stopHandlerBuilder_ = std::make_unique<StopHandlerBuilder>(*stopHandlerBuilder_);
425 newRunner.inputHolder_ = inputHolder_;
427 threadMpiMdrunnerAccessBarrier();
432 /*! \brief The callback used for running on spawned threads.
434 * Obtains the pointer to the master mdrunner object from the one
435 * argument permitted to the thread-launch API call, copies it to make
436 * a new runner for this thread, reinitializes necessary data, and
437 * proceeds to the simulation. */
438 static void mdrunner_start_fn(const void* arg)
442 const auto* masterMdrunner = reinterpret_cast<const gmx::Mdrunner*>(arg);
443 /* copy the arg list to make sure that it's thread-local. This
444 doesn't copy pointed-to items, of course; fnm, cr and fplog
445 are reset in the call below, all others should be const. */
446 gmx::Mdrunner mdrunner = masterMdrunner->cloneOnSpawnedThread();
449 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
453 void Mdrunner::spawnThreads(int numThreadsToLaunch)
456 /* now spawn new threads that start mdrunner_start_fn(), while
457 the main thread returns. Thread affinity is handled later. */
458 if (tMPI_Init_fn(TRUE, numThreadsToLaunch, TMPI_AFFINITY_NONE, mdrunner_start_fn, static_cast<const void*>(this))
461 GMX_THROW(gmx::InternalError("Failed to spawn thread-MPI threads"));
464 // Give the master thread the newly created valid communicator for
466 libraryWorldCommunicator = MPI_COMM_WORLD;
467 simulationCommunicator = MPI_COMM_WORLD;
468 threadMpiMdrunnerAccessBarrier();
470 GMX_UNUSED_VALUE(numThreadsToLaunch);
471 GMX_UNUSED_VALUE(mdrunner_start_fn);
477 /*! \brief Initialize variables for Verlet scheme simulation */
478 static void prepare_verlet_scheme(FILE* fplog,
482 const gmx_mtop_t& mtop,
484 bool makeGpuPairList,
485 const gmx::CpuInfo& cpuinfo)
487 // We checked the cut-offs in grompp, but double-check here.
488 // We have PME+LJcutoff kernels for rcoulomb>rvdw.
489 if (EEL_PME_EWALD(ir->coulombtype) && ir->vdwtype == VanDerWaalsType::Cut)
491 GMX_RELEASE_ASSERT(ir->rcoulomb >= ir->rvdw,
492 "With Verlet lists and PME we should have rcoulomb>=rvdw");
496 GMX_RELEASE_ASSERT(ir->rcoulomb == ir->rvdw,
497 "With Verlet lists and no PME rcoulomb and rvdw should be identical");
499 /* For NVE simulations, we will retain the initial list buffer */
500 if (EI_DYNAMICS(ir->eI) && ir->verletbuf_tol > 0
501 && !(EI_MD(ir->eI) && ir->etc == TemperatureCoupling::No))
503 /* Update the Verlet buffer size for the current run setup */
505 /* Here we assume SIMD-enabled kernels are being used. But as currently
506 * calc_verlet_buffer_size gives the same results for 4x8 and 4x4
507 * and 4x2 gives a larger buffer than 4x4, this is ok.
509 ListSetupType listType =
510 (makeGpuPairList ? ListSetupType::Gpu : ListSetupType::CpuSimdWhenSupported);
511 VerletbufListSetup listSetup = verletbufGetSafeListSetup(listType);
513 const real rlist_new =
514 calcVerletBufferSize(mtop, det(box), *ir, ir->nstlist, ir->nstlist - 1, -1, listSetup);
516 if (rlist_new != ir->rlist)
518 if (fplog != nullptr)
521 "\nChanging rlist from %g to %g for non-bonded %dx%d atom kernels\n\n",
524 listSetup.cluster_size_i,
525 listSetup.cluster_size_j);
527 ir->rlist = rlist_new;
531 if (nstlist_cmdline > 0 && (!EI_DYNAMICS(ir->eI) || ir->verletbuf_tol <= 0))
534 "Can not set nstlist without %s",
535 !EI_DYNAMICS(ir->eI) ? "dynamics" : "verlet-buffer-tolerance");
538 if (EI_DYNAMICS(ir->eI))
540 /* Set or try nstlist values */
541 increaseNstlist(fplog, cr, ir, nstlist_cmdline, &mtop, box, makeGpuPairList, cpuinfo);
545 /*! \brief Override the nslist value in inputrec
547 * with value passed on the command line (if any)
549 static void override_nsteps_cmdline(const gmx::MDLogger& mdlog, int64_t nsteps_cmdline, t_inputrec* ir)
553 /* override with anything else than the default -2 */
554 if (nsteps_cmdline > -2)
556 char sbuf_steps[STEPSTRSIZE];
557 char sbuf_msg[STRLEN];
559 ir->nsteps = nsteps_cmdline;
560 if (EI_DYNAMICS(ir->eI) && nsteps_cmdline != -1)
563 "Overriding nsteps with value passed on the command line: %s steps, %.3g ps",
564 gmx_step_str(nsteps_cmdline, sbuf_steps),
565 fabs(nsteps_cmdline * ir->delta_t));
570 "Overriding nsteps with value passed on the command line: %s steps",
571 gmx_step_str(nsteps_cmdline, sbuf_steps));
574 GMX_LOG(mdlog.warning).asParagraph().appendText(sbuf_msg);
576 else if (nsteps_cmdline < -2)
578 gmx_fatal(FARGS, "Invalid nsteps value passed on the command line: %" PRId64, nsteps_cmdline);
580 /* Do nothing if nsteps_cmdline == -2 */
586 /*! \brief Return whether GPU acceleration of nonbondeds is supported with the given settings.
588 * If not, and if a warning may be issued, logs a warning about
589 * falling back to CPU code. With thread-MPI, only the first
590 * call to this function should have \c issueWarning true. */
591 static bool gpuAccelerationOfNonbondedIsUseful(const MDLogger& mdlog, const t_inputrec& ir, bool issueWarning)
593 bool gpuIsUseful = true;
596 if (ir.opts.ngener - ir.nwall > 1)
598 /* The GPU code does not support more than one energy group.
599 * If the user requested GPUs explicitly, a fatal error is given later.
603 "Multiple energy groups is not implemented for GPUs, falling back to the CPU. "
604 "For better performance, run on the GPU without energy groups and then do "
605 "gmx mdrun -rerun option on the trajectory with an energy group .tpr file.";
611 warning = "TPI is not implemented for GPUs.";
614 if (!gpuIsUseful && issueWarning)
616 GMX_LOG(mdlog.warning).asParagraph().appendText(warning);
622 //! Initializes the logger for mdrun.
623 static gmx::LoggerOwner buildLogger(FILE* fplog, const bool isSimulationMasterRank)
625 gmx::LoggerBuilder builder;
626 if (fplog != nullptr)
628 builder.addTargetFile(gmx::MDLogger::LogLevel::Info, fplog);
630 if (isSimulationMasterRank)
632 builder.addTargetStream(gmx::MDLogger::LogLevel::Warning, &gmx::TextOutputFile::standardError());
634 return builder.build();
637 //! Make a TaskTarget from an mdrun argument string.
638 static TaskTarget findTaskTarget(const char* optionString)
640 TaskTarget returnValue = TaskTarget::Auto;
642 if (strncmp(optionString, "auto", 3) == 0)
644 returnValue = TaskTarget::Auto;
646 else if (strncmp(optionString, "cpu", 3) == 0)
648 returnValue = TaskTarget::Cpu;
650 else if (strncmp(optionString, "gpu", 3) == 0)
652 returnValue = TaskTarget::Gpu;
656 GMX_ASSERT(false, "Option string should have been checked for sanity already");
662 //! Finish run, aggregate data to print performance info.
663 static void finish_run(FILE* fplog,
664 const gmx::MDLogger& mdlog,
666 const t_inputrec& inputrec,
668 gmx_wallcycle* wcycle,
669 gmx_walltime_accounting_t walltime_accounting,
670 nonbonded_verlet_t* nbv,
671 const gmx_pme_t* pme,
675 double nbfs = 0, mflop = 0;
676 double elapsed_time, elapsed_time_over_all_ranks, elapsed_time_over_all_threads,
677 elapsed_time_over_all_threads_over_all_ranks;
678 /* Control whether it is valid to print a report. Only the
679 simulation master may print, but it should not do so if the run
680 terminated e.g. before a scheduled reset step. This is
681 complicated by the fact that PME ranks are unaware of the
682 reason why they were sent a pmerecvqxFINISH. To avoid
683 communication deadlocks, we always do the communication for the
684 report, even if we've decided not to write the report, because
685 how long it takes to finish the run is not important when we've
686 decided not to report on the simulation performance.
688 Further, we only report performance for dynamical integrators,
689 because those are the only ones for which we plan to
690 consider doing any optimizations. */
691 bool printReport = EI_DYNAMICS(inputrec.eI) && SIMMASTER(cr);
693 if (printReport && !walltime_accounting_get_valid_finish(walltime_accounting))
695 GMX_LOG(mdlog.warning)
697 .appendText("Simulation ended prematurely, no performance report will be written.");
702 std::unique_ptr<t_nrnb> nrnbTotalStorage;
705 nrnbTotalStorage = std::make_unique<t_nrnb>();
706 nrnb_tot = nrnbTotalStorage.get();
708 MPI_Allreduce(nrnb->n.data(), nrnb_tot->n.data(), eNRNB, MPI_DOUBLE, MPI_SUM, cr->mpi_comm_mysim);
716 elapsed_time = walltime_accounting_get_time_since_reset(walltime_accounting);
717 elapsed_time_over_all_threads =
718 walltime_accounting_get_time_since_reset_over_all_threads(walltime_accounting);
722 /* reduce elapsed_time over all MPI ranks in the current simulation */
723 MPI_Allreduce(&elapsed_time, &elapsed_time_over_all_ranks, 1, MPI_DOUBLE, MPI_SUM, cr->mpi_comm_mysim);
724 elapsed_time_over_all_ranks /= cr->nnodes;
725 /* Reduce elapsed_time_over_all_threads over all MPI ranks in the
726 * current simulation. */
727 MPI_Allreduce(&elapsed_time_over_all_threads,
728 &elapsed_time_over_all_threads_over_all_ranks,
737 elapsed_time_over_all_ranks = elapsed_time;
738 elapsed_time_over_all_threads_over_all_ranks = elapsed_time_over_all_threads;
743 print_flop(fplog, nrnb_tot, &nbfs, &mflop);
746 if (thisRankHasDuty(cr, DUTY_PP) && DOMAINDECOMP(cr))
748 print_dd_statistics(cr, inputrec, fplog);
751 /* TODO Move the responsibility for any scaling by thread counts
752 * to the code that handled the thread region, so that there's a
753 * mechanism to keep cycle counting working during the transition
754 * to task parallelism. */
755 int nthreads_pp = gmx_omp_nthreads_get(ModuleMultiThread::Nonbonded);
756 int nthreads_pme = gmx_omp_nthreads_get(ModuleMultiThread::Pme);
757 wallcycle_scale_by_num_threads(
758 wcycle, thisRankHasDuty(cr, DUTY_PME) && !thisRankHasDuty(cr, DUTY_PP), nthreads_pp, nthreads_pme);
759 auto cycle_sum(wallcycle_sum(cr, wcycle));
763 auto* nbnxn_gpu_timings =
764 (nbv != nullptr && nbv->useGpu()) ? Nbnxm::gpu_get_timings(nbv->gpu_nbv) : nullptr;
765 gmx_wallclock_gpu_pme_t pme_gpu_timings = {};
767 if (pme_gpu_task_enabled(pme))
769 pme_gpu_get_timings(pme, &pme_gpu_timings);
771 wallcycle_print(fplog,
777 elapsed_time_over_all_ranks,
783 if (EI_DYNAMICS(inputrec.eI))
785 delta_t = inputrec.delta_t;
791 elapsed_time_over_all_threads_over_all_ranks,
792 elapsed_time_over_all_ranks,
793 walltime_accounting_get_nsteps_done_since_reset(walltime_accounting),
801 elapsed_time_over_all_threads_over_all_ranks,
802 elapsed_time_over_all_ranks,
803 walltime_accounting_get_nsteps_done_since_reset(walltime_accounting),
811 int Mdrunner::mdrunner()
814 std::unique_ptr<t_forcerec> fr;
815 real ewaldcoeff_q = 0;
816 real ewaldcoeff_lj = 0;
817 int nChargePerturbed = -1, nTypePerturbed = 0;
818 gmx_walltime_accounting_t walltime_accounting = nullptr;
819 MembedHolder membedHolder(filenames.size(), filenames.data());
821 /* CAUTION: threads may be started later on in this function, so
822 cr doesn't reflect the final parallel state right now */
825 /* TODO: inputrec should tell us whether we use an algorithm, not a file option */
826 const bool doEssentialDynamics = opt2bSet("-ei", filenames.size(), filenames.data());
827 const bool doRerun = mdrunOptions.rerun;
829 // Handle task-assignment related user options.
830 EmulateGpuNonbonded emulateGpuNonbonded =
831 (getenv("GMX_EMULATE_GPU") != nullptr ? EmulateGpuNonbonded::Yes : EmulateGpuNonbonded::No);
833 std::vector<int> userGpuTaskAssignment;
836 userGpuTaskAssignment = parseUserTaskAssignmentString(hw_opt.userGpuTaskAssignment);
838 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
839 auto nonbondedTarget = findTaskTarget(nbpu_opt);
840 auto pmeTarget = findTaskTarget(pme_opt);
841 auto pmeFftTarget = findTaskTarget(pme_fft_opt);
842 auto bondedTarget = findTaskTarget(bonded_opt);
843 auto updateTarget = findTaskTarget(update_opt);
845 FILE* fplog = nullptr;
846 // If we are appending, we don't write log output because we need
847 // to check that the old log file matches what the checkpoint file
848 // expects. Otherwise, we should start to write log output now if
849 // there is a file ready for it.
850 if (logFileHandle != nullptr && startingBehavior != StartingBehavior::RestartWithAppending)
852 fplog = gmx_fio_getfp(logFileHandle);
854 const bool isSimulationMasterRank = findIsSimulationMasterRank(ms, simulationCommunicator);
855 gmx::LoggerOwner logOwner(buildLogger(fplog, isSimulationMasterRank));
856 gmx::MDLogger mdlog(logOwner.logger());
858 gmx_print_detected_hardware(fplog, isSimulationMasterRank && isMasterSim(ms), mdlog, hwinfo_);
860 std::vector<int> availableDevices =
861 makeListOfAvailableDevices(hwinfo_->deviceInfoList, hw_opt.devicesSelectedByUser);
862 const int numAvailableDevices = gmx::ssize(availableDevices);
864 // Print citation requests after all software/hardware printing
865 pleaseCiteGromacs(fplog);
867 // Note: legacy program logic relies on checking whether these pointers are assigned.
868 // Objects may or may not be allocated later.
869 std::unique_ptr<t_inputrec> inputrec;
870 std::unique_ptr<t_state> globalState;
872 auto partialDeserializedTpr = std::make_unique<PartialDeserializedTprFile>();
874 if (isSimulationMasterRank)
876 // Allocate objects to be initialized by later function calls.
877 /* Only the master rank has the global state */
878 globalState = std::make_unique<t_state>();
879 inputrec = std::make_unique<t_inputrec>();
881 /* Read (nearly) all data required for the simulation
882 * and keep the partly serialized tpr contents to send to other ranks later
884 applyGlobalSimulationState(
885 *inputHolder_.get(), partialDeserializedTpr.get(), globalState.get(), inputrec.get(), &mtop);
888 /* Check and update the hardware options for internal consistency */
889 checkAndUpdateHardwareOptions(
890 mdlog, &hw_opt, isSimulationMasterRank, domdecOptions.numPmeRanks, inputrec.get());
892 if (GMX_THREAD_MPI && isSimulationMasterRank)
894 bool useGpuForNonbonded = false;
895 bool useGpuForPme = false;
898 GMX_RELEASE_ASSERT(inputrec != nullptr, "Keep the compiler happy");
900 // If the user specified the number of ranks, then we must
901 // respect that, but in default mode, we need to allow for
902 // the number of GPUs to choose the number of ranks.
903 auto canUseGpuForNonbonded = buildSupportsNonbondedOnGpu(nullptr);
904 useGpuForNonbonded = decideWhetherToUseGpusForNonbondedWithThreadMpi(
906 numAvailableDevices > 0,
907 userGpuTaskAssignment,
909 canUseGpuForNonbonded,
910 gpuAccelerationOfNonbondedIsUseful(mdlog, *inputrec, GMX_THREAD_MPI),
911 hw_opt.nthreads_tmpi);
912 useGpuForPme = decideWhetherToUseGpusForPmeWithThreadMpi(useGpuForNonbonded,
915 userGpuTaskAssignment,
918 hw_opt.nthreads_tmpi,
919 domdecOptions.numPmeRanks);
921 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
923 /* Determine how many thread-MPI ranks to start.
925 * TODO Over-writing the user-supplied value here does
926 * prevent any possible subsequent checks from working
928 hw_opt.nthreads_tmpi = get_nthreads_mpi(hwinfo_,
936 membedHolder.doMembed());
938 // Now start the threads for thread MPI.
939 spawnThreads(hw_opt.nthreads_tmpi);
940 // The spawned threads enter mdrunner() and execution of
941 // master and spawned threads joins at the end of this block.
944 GMX_RELEASE_ASSERT(ms || simulationCommunicator != MPI_COMM_NULL,
945 "Must have valid communicator unless running a multi-simulation");
946 CommrecHandle crHandle = init_commrec(simulationCommunicator);
947 t_commrec* cr = crHandle.get();
948 GMX_RELEASE_ASSERT(cr != nullptr, "Must have valid commrec");
950 PhysicalNodeCommunicator physicalNodeComm(libraryWorldCommunicator, gmx_physicalnode_id_hash());
952 // If we detected the topology on this system, double-check that it makes sense
953 if (hwinfo_->hardwareTopology->isThisSystem())
955 hardwareTopologyDoubleCheckDetection(mdlog, *hwinfo_->hardwareTopology);
960 /* now broadcast everything to the non-master nodes/threads: */
961 if (!isSimulationMasterRank)
963 // Until now, only the master rank has a non-null pointer.
964 // On non-master ranks, allocate the object that will receive data in the following call.
965 inputrec = std::make_unique<t_inputrec>();
967 init_parallel(cr->mpiDefaultCommunicator,
971 partialDeserializedTpr.get());
973 GMX_RELEASE_ASSERT(inputrec != nullptr, "All ranks should have a valid inputrec now");
974 partialDeserializedTpr.reset(nullptr);
977 !inputrec->useConstantAcceleration,
978 "Linear acceleration has been removed in GROMACS 2022, and was broken for many years "
979 "before that. Use GROMACS 4.5 or earlier if you need this feature.");
981 // Now the number of ranks is known to all ranks, and each knows
982 // the inputrec read by the master rank. The ranks can now all run
983 // the task-deciding functions and will agree on the result
984 // without needing to communicate.
985 const bool useDomainDecomposition =
986 (PAR(cr) && !(EI_TPI(inputrec->eI) || inputrec->eI == IntegrationAlgorithm::NM));
988 // Note that these variables describe only their own node.
990 // Note that when bonded interactions run on a GPU they always run
991 // alongside a nonbonded task, so do not influence task assignment
992 // even though they affect the force calculation workload.
993 bool useGpuForNonbonded = false;
994 bool useGpuForPme = false;
995 bool useGpuForBonded = false;
996 bool useGpuForUpdate = false;
997 bool gpusWereDetected = hwinfo_->ngpu_compatible_tot > 0;
1000 // It's possible that there are different numbers of GPUs on
1001 // different nodes, which is the user's responsibility to
1002 // handle. If unsuitable, we will notice that during task
1004 auto canUseGpuForNonbonded = buildSupportsNonbondedOnGpu(nullptr);
1005 useGpuForNonbonded = decideWhetherToUseGpusForNonbonded(
1007 userGpuTaskAssignment,
1008 emulateGpuNonbonded,
1009 canUseGpuForNonbonded,
1010 gpuAccelerationOfNonbondedIsUseful(mdlog, *inputrec, !GMX_THREAD_MPI),
1012 useGpuForPme = decideWhetherToUseGpusForPme(useGpuForNonbonded,
1014 userGpuTaskAssignment,
1017 cr->sizeOfDefaultCommunicator,
1018 domdecOptions.numPmeRanks,
1020 useGpuForBonded = decideWhetherToUseGpusForBonded(
1021 useGpuForNonbonded, useGpuForPme, bondedTarget, *inputrec, mtop, domdecOptions.numPmeRanks, gpusWereDetected);
1023 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1025 const PmeRunMode pmeRunMode = determinePmeRunMode(useGpuForPme, pmeFftTarget, *inputrec);
1027 // Initialize development feature flags that enabled by environment variable
1028 // and report those features that are enabled.
1029 const DevelopmentFeatureFlags devFlags =
1030 manageDevelopmentFeatures(mdlog, useGpuForNonbonded, pmeRunMode);
1032 const bool useModularSimulator = checkUseModularSimulator(false,
1039 doEssentialDynamics,
1040 membedHolder.doMembed());
1042 // Build restraints.
1043 // TODO: hide restraint implementation details from Mdrunner.
1044 // There is nothing unique about restraints at this point as far as the
1045 // Mdrunner is concerned. The Mdrunner should just be getting a sequence of
1046 // factory functions from the SimulationContext on which to call mdModules_->add().
1047 // TODO: capture all restraints into a single RestraintModule, passed to the runner builder.
1048 for (auto&& restraint : restraintManager_->getRestraints())
1050 auto module = RestraintMDModule::create(restraint, restraint->sites());
1051 mdModules_->add(std::move(module));
1054 // TODO: Error handling
1055 mdModules_->assignOptionsToModules(*inputrec->params, nullptr);
1056 // now that the MDModules know their options, they know which callbacks to sign up to
1057 mdModules_->subscribeToSimulationSetupNotifications();
1058 const auto& setupNotifier = mdModules_->notifiers().simulationSetupNotifier_;
1060 if (inputrec->internalParameters != nullptr)
1062 setupNotifier.notify(*inputrec->internalParameters);
1065 if (fplog != nullptr)
1067 pr_inputrec(fplog, 0, "Input Parameters", inputrec.get(), FALSE);
1068 fprintf(fplog, "\n");
1073 /* In rerun, set velocities to zero if present */
1074 if (doRerun && ((globalState->flags & enumValueToBitMask(StateEntry::V)) != 0))
1076 // rerun does not use velocities
1080 "Rerun trajectory contains velocities. Rerun does only evaluate "
1081 "potential energy and forces. The velocities will be ignored.");
1082 for (int i = 0; i < globalState->natoms; i++)
1084 clear_rvec(globalState->v[i]);
1086 globalState->flags &= ~enumValueToBitMask(StateEntry::V);
1089 /* now make sure the state is initialized and propagated */
1090 set_state_entries(globalState.get(), inputrec.get(), useModularSimulator);
1093 /* NM and TPI parallelize over force/energy calculations, not atoms,
1094 * so we need to initialize and broadcast the global state.
1096 if (inputrec->eI == IntegrationAlgorithm::NM || inputrec->eI == IntegrationAlgorithm::TPI)
1100 globalState = std::make_unique<t_state>();
1102 broadcastStateWithoutDynamics(
1103 cr->mpiDefaultCommunicator, DOMAINDECOMP(cr), PAR(cr), globalState.get());
1106 /* A parallel command line option consistency check that we can
1107 only do after any threads have started. */
1109 && (domdecOptions.numCells[XX] > 1 || domdecOptions.numCells[YY] > 1
1110 || domdecOptions.numCells[ZZ] > 1 || domdecOptions.numPmeRanks > 0))
1113 "The -dd or -npme option request a parallel simulation, "
1115 "but %s was compiled without threads or MPI enabled",
1116 output_env_get_program_display_name(oenv));
1117 #elif GMX_THREAD_MPI
1118 "but the number of MPI-threads (option -ntmpi) is not set or is 1");
1120 "but %s was not started through mpirun/mpiexec or only one rank was requested "
1121 "through mpirun/mpiexec",
1122 output_env_get_program_display_name(oenv));
1126 if (doRerun && (EI_ENERGY_MINIMIZATION(inputrec->eI) || IntegrationAlgorithm::NM == inputrec->eI))
1129 "The .mdp file specified an energy mininization or normal mode algorithm, and "
1130 "these are not compatible with mdrun -rerun");
1133 /* NMR restraints must be initialized before load_checkpoint,
1134 * since with time averaging the history is added to t_state.
1135 * For proper consistency check we therefore need to extend
1137 * So the PME-only nodes (if present) will also initialize
1138 * the distance restraints.
1141 /* This needs to be called before read_checkpoint to extend the state */
1142 t_disresdata* disresdata;
1143 snew(disresdata, 1);
1147 DisResRunMode::MDRun,
1148 MASTER(cr) ? DDRole::Master : DDRole::Agent,
1149 PAR(cr) ? NumRanks::Multiple : NumRanks::Single,
1154 replExParams.exchangeInterval > 0);
1156 std::unique_ptr<t_oriresdata> oriresData;
1157 if (gmx_mtop_ftype_count(mtop, F_ORIRES) > 0)
1159 oriresData = std::make_unique<t_oriresdata>(fplog, mtop, *inputrec, cr, ms, globalState.get());
1162 auto deform = prepareBoxDeformation(globalState != nullptr ? globalState->box : box,
1163 MASTER(cr) ? DDRole::Master : DDRole::Agent,
1164 PAR(cr) ? NumRanks::Multiple : NumRanks::Single,
1165 cr->mpi_comm_mygroup,
1169 /* We have to remember the generation's first step before reading checkpoint.
1170 This way, we can report to the F@H core both the generation's first step
1171 and the restored first step, thus making it able to distinguish between
1172 an interruption/resume and start of the n-th generation simulation.
1173 Having this information, the F@H core can correctly calculate and report
1176 int gen_first_step = 0;
1179 gen_first_step = inputrec->init_step;
1183 ObservablesHistory observablesHistory = {};
1185 auto modularSimulatorCheckpointData = std::make_unique<ReadCheckpointDataHolder>();
1186 if (startingBehavior != StartingBehavior::NewSimulation)
1188 /* Check if checkpoint file exists before doing continuation.
1189 * This way we can use identical input options for the first and subsequent runs...
1191 if (mdrunOptions.numStepsCommandline > -2)
1193 /* Temporarily set the number of steps to unlimited to avoid
1194 * triggering the nsteps check in load_checkpoint().
1195 * This hack will go away soon when the -nsteps option is removed.
1197 inputrec->nsteps = -1;
1200 // Finish applying initial simulation state information from external sources on all ranks.
1201 // Reconcile checkpoint file data with Mdrunner state established up to this point.
1202 applyLocalState(*inputHolder_.get(),
1205 domdecOptions.numCells,
1208 &observablesHistory,
1209 mdrunOptions.reproducible,
1210 mdModules_->notifiers(),
1211 modularSimulatorCheckpointData.get(),
1212 useModularSimulator);
1213 // TODO: (#3652) Synchronize filesystem state, SimulationInput contents, and program
1215 // on all code paths.
1216 // Write checkpoint or provide hook to update SimulationInput.
1217 // If there was a checkpoint file, SimulationInput contains more information
1218 // than if there wasn't. At this point, we have synchronized the in-memory
1219 // state with the filesystem state only for restarted simulations. We should
1220 // be calling applyLocalState unconditionally and expect that the completeness
1221 // of SimulationInput is not dependent on its creation method.
1223 if (startingBehavior == StartingBehavior::RestartWithAppending && logFileHandle)
1225 // Now we can start normal logging to the truncated log file.
1226 fplog = gmx_fio_getfp(logFileHandle);
1227 prepareLogAppending(fplog);
1228 logOwner = buildLogger(fplog, MASTER(cr));
1229 mdlog = logOwner.logger();
1236 fcRegisterSteps(inputrec->nsteps + inputrec->init_step, gen_first_step);
1240 if (mdrunOptions.numStepsCommandline > -2)
1245 "The -nsteps functionality is deprecated, and may be removed in a future "
1247 "Consider using gmx convert-tpr -nsteps or changing the appropriate .mdp "
1250 /* override nsteps with value set on the commandline */
1251 override_nsteps_cmdline(mdlog, mdrunOptions.numStepsCommandline, inputrec.get());
1253 if (isSimulationMasterRank)
1255 copy_mat(globalState->box, box);
1260 gmx_bcast(sizeof(box), box, cr->mpiDefaultCommunicator);
1263 if (inputrec->cutoff_scheme != CutoffScheme::Verlet)
1266 "This group-scheme .tpr file can no longer be run by mdrun. Please update to the "
1267 "Verlet scheme, or use an earlier version of GROMACS if necessary.");
1269 /* Update rlist and nstlist. */
1270 /* Note: prepare_verlet_scheme is calling increaseNstlist(...), which (while attempting to
1271 * increase rlist) tries to check if the newly chosen value fits with the DD scheme. As this is
1272 * run before any DD scheme is set up, this check is never executed. See #3334 for more details.
1274 prepare_verlet_scheme(fplog,
1280 useGpuForNonbonded || (emulateGpuNonbonded == EmulateGpuNonbonded::Yes),
1283 // We need to decide on update groups early, as this affects
1284 // inter-domain communication distances.
1285 auto updateGroupingsPerMoleculeType = makeUpdateGroupingsPerMoleculeType(mtop);
1286 const real maxUpdateGroupRadius = computeMaxUpdateGroupRadius(
1287 mtop, updateGroupingsPerMoleculeType, maxReferenceTemperature(*inputrec));
1288 const real cutoffMargin = std::sqrt(max_cutoff2(inputrec->pbcType, box)) - inputrec->rlist;
1289 UpdateGroups updateGroups = makeUpdateGroups(mdlog,
1290 std::move(updateGroupingsPerMoleculeType),
1291 maxUpdateGroupRadius,
1292 useDomainDecomposition,
1293 systemHasConstraintsOrVsites(mtop),
1296 // This builder is necessary while we have multi-part construction
1297 // of DD. Before DD is constructed, we use the existence of
1298 // the builder object to indicate that further construction of DD
1300 std::unique_ptr<DomainDecompositionBuilder> ddBuilder;
1301 if (useDomainDecomposition)
1303 ddBuilder = std::make_unique<DomainDecompositionBuilder>(
1310 mdModules_->notifiers(),
1312 updateGroups.updateGroupingPerMoleculeType(),
1313 updateGroups.useUpdateGroups(),
1314 updateGroups.maxUpdateGroupRadius(),
1315 positionsFromStatePointer(globalState.get()),
1321 /* PME, if used, is done on all nodes with 1D decomposition */
1322 cr->nnodes = cr->sizeOfDefaultCommunicator;
1323 cr->sim_nodeid = cr->rankInDefaultCommunicator;
1324 cr->nodeid = cr->rankInDefaultCommunicator;
1326 cr->duty = (DUTY_PP | DUTY_PME);
1328 if (inputrec->pbcType == PbcType::Screw)
1330 gmx_fatal(FARGS, "pbc=screw is only implemented with domain decomposition");
1334 // Produce the task assignment for this rank - done after DD is constructed
1335 GpuTaskAssignments gpuTaskAssignments = GpuTaskAssignmentsBuilder::build(
1337 userGpuTaskAssignment,
1339 simulationCommunicator,
1347 thisRankHasDuty(cr, DUTY_PP),
1348 // TODO cr->duty & DUTY_PME should imply that a PME
1349 // algorithm is active, but currently does not.
1350 EEL_PME(inputrec->coulombtype) && thisRankHasDuty(cr, DUTY_PME));
1352 // Get the device handles for the modules, nullptr when no task is assigned.
1354 DeviceInformation* deviceInfo = gpuTaskAssignments.initDevice(&deviceId);
1356 // timing enabling - TODO put this in gpu_utils (even though generally this is just option handling?)
1357 bool useTiming = true;
1361 /* WARNING: CUDA timings are incorrect with multiple streams.
1362 * This is the main reason why they are disabled by default.
1364 // TODO: Consider turning on by default when we can detect nr of streams.
1365 useTiming = (getenv("GMX_ENABLE_GPU_TIMING") != nullptr);
1367 else if (GMX_GPU_OPENCL)
1369 useTiming = (getenv("GMX_DISABLE_GPU_TIMING") == nullptr);
1372 // TODO Currently this is always built, yet DD partition code
1373 // checks if it is built before using it. Probably it should
1374 // become an MDModule that is made only when another module
1375 // requires it (e.g. pull, CompEl, density fitting), so that we
1376 // don't update the local atom sets unilaterally every step.
1377 LocalAtomSetManager atomSets;
1380 // TODO Pass the GPU streams to ddBuilder to use in buffer
1381 // transfers (e.g. halo exchange)
1382 cr->dd = ddBuilder->build(&atomSets);
1383 // The builder's job is done, so destruct it
1384 ddBuilder.reset(nullptr);
1385 // Note that local state still does not exist yet.
1388 // The GPU update is decided here because we need to know whether the constraints or
1389 // SETTLEs can span accross the domain borders (i.e. whether or not update groups are
1390 // defined). This is only known after DD is initialized, hence decision on using GPU
1391 // update is done so late.
1394 const bool haveFrozenAtoms = inputrecFrozenAtoms(inputrec.get());
1396 useGpuForUpdate = decideWhetherToUseGpuForUpdate(useDomainDecomposition,
1397 updateGroups.useUpdateGroups(),
1399 domdecOptions.numPmeRanks > 0,
1405 doEssentialDynamics,
1406 gmx_mtop_ftype_count(mtop, F_ORIRES) > 0,
1412 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1414 const bool printHostName = (cr->nnodes > 1);
1415 gpuTaskAssignments.reportGpuUsage(mdlog, printHostName, useGpuForBonded, pmeRunMode, useGpuForUpdate);
1417 const bool disableNonbondedCalculation = (getenv("GMX_NO_NONBONDED") != nullptr);
1418 if (disableNonbondedCalculation)
1420 /* turn off non-bonded calculations */
1421 GMX_LOG(mdlog.warning)
1424 "Found environment variable GMX_NO_NONBONDED.\n"
1425 "Disabling nonbonded calculations.");
1428 MdrunScheduleWorkload runScheduleWork;
1430 bool useGpuDirectHalo = decideWhetherToUseGpuForHalo(devFlags,
1431 havePPDomainDecomposition(cr),
1433 useModularSimulator,
1435 EI_ENERGY_MINIMIZATION(inputrec->eI));
1437 // Also populates the simulation constant workload description.
1438 runScheduleWork.simulationWork = createSimulationWorkload(*inputrec,
1439 disableNonbondedCalculation,
1447 std::unique_ptr<DeviceStreamManager> deviceStreamManager = nullptr;
1449 if (deviceInfo != nullptr)
1451 if (DOMAINDECOMP(cr) && thisRankHasDuty(cr, DUTY_PP))
1453 dd_setup_dlb_resource_sharing(cr, deviceId);
1455 deviceStreamManager = std::make_unique<DeviceStreamManager>(
1456 *deviceInfo, havePPDomainDecomposition(cr), runScheduleWork.simulationWork, useTiming);
1459 // If the user chose a task assignment, give them some hints
1460 // where appropriate.
1461 if (!userGpuTaskAssignment.empty())
1463 gpuTaskAssignments.logPerformanceHints(mdlog, numAvailableDevices);
1468 /* After possible communicator splitting in make_dd_communicators.
1469 * we can set up the intra/inter node communication.
1471 gmx_setup_nodecomm(fplog, cr);
1477 GMX_LOG(mdlog.warning)
1479 .appendTextFormatted(
1480 "This is simulation %d out of %d running as a composite GROMACS\n"
1481 "multi-simulation job. Setup for this simulation:\n",
1482 ms->simulationIndex_,
1483 ms->numSimulations_);
1485 GMX_LOG(mdlog.warning)
1486 .appendTextFormatted("Using %d MPI %s\n",
1489 cr->nnodes == 1 ? "thread" : "threads"
1491 cr->nnodes == 1 ? "process" : "processes"
1497 // If mdrun -pin auto honors any affinity setting that already
1498 // exists. If so, it is nice to provide feedback about whether
1499 // that existing affinity setting was from OpenMP or something
1500 // else, so we run this code both before and after we initialize
1501 // the OpenMP support.
1502 gmx_check_thread_affinity_set(mdlog, &hw_opt, hwinfo_->nthreads_hw_avail, FALSE);
1503 /* Check and update the number of OpenMP threads requested */
1504 checkAndUpdateRequestedNumOpenmpThreads(
1505 &hw_opt, *hwinfo_, cr, ms, physicalNodeComm.size_, pmeRunMode, mtop, *inputrec);
1507 gmx_omp_nthreads_init(mdlog,
1509 hwinfo_->nthreads_hw_avail,
1510 physicalNodeComm.size_,
1511 hw_opt.nthreads_omp,
1512 hw_opt.nthreads_omp_pme,
1513 !thisRankHasDuty(cr, DUTY_PP));
1515 const bool bEnableFPE = gmxShouldEnableFPExceptions();
1516 // FIXME - reconcile with gmx_feenableexcept() call from CommandLineModuleManager::run()
1519 gmx_feenableexcept();
1522 /* Now that we know the setup is consistent, check for efficiency */
1523 check_resource_division_efficiency(
1524 hwinfo_, gpuTaskAssignments.thisRankHasAnyGpuTask(), mdrunOptions.ntompOptionIsSet, cr, mdlog);
1526 /* getting number of PP/PME threads on this MPI / tMPI rank.
1527 PME: env variable should be read only on one node to make sure it is
1528 identical everywhere;
1530 const int numThreadsOnThisRank = thisRankHasDuty(cr, DUTY_PP)
1531 ? gmx_omp_nthreads_get(ModuleMultiThread::Nonbonded)
1532 : gmx_omp_nthreads_get(ModuleMultiThread::Pme);
1533 checkHardwareOversubscription(
1534 numThreadsOnThisRank, cr->nodeid, *hwinfo_->hardwareTopology, physicalNodeComm, mdlog);
1536 // Enable Peer access between GPUs where available
1537 // Only for DD, only master PP rank needs to perform setup, and only if thread MPI plus
1538 // any of the GPU communication features are active.
1539 if (DOMAINDECOMP(cr) && MASTER(cr) && thisRankHasDuty(cr, DUTY_PP) && GMX_THREAD_MPI
1540 && (runScheduleWork.simulationWork.useGpuHaloExchange
1541 || runScheduleWork.simulationWork.useGpuPmePpCommunication))
1543 setupGpuDevicePeerAccess(gpuTaskAssignments.deviceIdsAssigned(), mdlog);
1546 if (hw_opt.threadAffinity != ThreadAffinity::Off)
1548 /* Before setting affinity, check whether the affinity has changed
1549 * - which indicates that probably the OpenMP library has changed it
1550 * since we first checked).
1552 gmx_check_thread_affinity_set(mdlog, &hw_opt, hwinfo_->nthreads_hw_avail, TRUE);
1554 int numThreadsOnThisNode, intraNodeThreadOffset;
1555 analyzeThreadsOnThisNode(
1556 physicalNodeComm, numThreadsOnThisRank, &numThreadsOnThisNode, &intraNodeThreadOffset);
1558 /* Set the CPU affinity */
1559 gmx_set_thread_affinity(mdlog,
1562 *hwinfo_->hardwareTopology,
1563 numThreadsOnThisRank,
1564 numThreadsOnThisNode,
1565 intraNodeThreadOffset,
1569 if (mdrunOptions.timingOptions.resetStep > -1)
1574 "The -resetstep functionality is deprecated, and may be removed in a "
1577 std::unique_ptr<gmx_wallcycle> wcycle =
1578 wallcycle_init(fplog, mdrunOptions.timingOptions.resetStep, cr);
1582 /* Master synchronizes its value of reset_counters with all nodes
1583 * including PME only nodes */
1584 int64_t reset_counters = wcycle_get_reset_counters(wcycle.get());
1585 gmx_bcast(sizeof(reset_counters), &reset_counters, cr->mpi_comm_mysim);
1586 wcycle_set_reset_counters(wcycle.get(), reset_counters);
1589 // Membrane embedding must be initialized before we call init_forcerec()
1590 membedHolder.initializeMembed(fplog,
1597 &mdrunOptions.checkpointOptions.period);
1599 const bool thisRankHasPmeGpuTask = gpuTaskAssignments.thisRankHasPmeGpuTask();
1600 std::unique_ptr<MDAtoms> mdAtoms;
1601 std::unique_ptr<VirtualSitesHandler> vsite;
1602 std::unique_ptr<ListedForcesGpu> listedForcesGpu;
1605 if (thisRankHasDuty(cr, DUTY_PP))
1607 setupNotifier.notify(*cr);
1608 setupNotifier.notify(&atomSets);
1609 setupNotifier.notify(mtop);
1610 setupNotifier.notify(inputrec->pbcType);
1611 setupNotifier.notify(SimulationTimeStep{ inputrec->delta_t });
1612 /* Initiate forcerecord */
1613 fr = std::make_unique<t_forcerec>();
1614 fr->forceProviders = mdModules_->initForceProviders();
1615 init_forcerec(fplog,
1622 opt2fn("-table", filenames.size(), filenames.data()),
1623 opt2fn("-tablep", filenames.size(), filenames.data()),
1624 opt2fns("-tableb", filenames.size(), filenames.data()),
1626 // Dirty hack, for fixing disres and orires should be made mdmodules
1627 fr->fcdata->disres = disresdata;
1628 fr->fcdata->orires.swap(oriresData);
1630 // Save a handle to device stream manager to use elsewhere in the code
1631 // TODO: Forcerec is not a correct place to store it.
1632 fr->deviceStreamManager = deviceStreamManager.get();
1634 if (runScheduleWork.simulationWork.useGpuPmePpCommunication && !thisRankHasDuty(cr, DUTY_PME))
1637 deviceStreamManager != nullptr,
1638 "GPU device stream manager should be valid in order to use PME-PP direct "
1641 deviceStreamManager->streamIsValid(DeviceStreamType::PmePpTransfer),
1642 "GPU PP-PME stream should be valid in order to use GPU PME-PP direct "
1644 fr->pmePpCommGpu = std::make_unique<gmx::PmePpCommGpu>(
1647 deviceStreamManager->context(),
1648 deviceStreamManager->stream(DeviceStreamType::PmePpTransfer));
1651 fr->nbv = Nbnxm::init_nb_verlet(mdlog,
1656 runScheduleWork.simulationWork.useGpuNonbonded,
1657 deviceStreamManager.get(),
1661 // TODO: Move the logic below to a GPU bonded builder
1662 if (runScheduleWork.simulationWork.useGpuBonded)
1664 GMX_RELEASE_ASSERT(deviceStreamManager != nullptr,
1665 "GPU device stream manager should be valid in order to use GPU "
1666 "version of bonded forces.");
1667 listedForcesGpu = std::make_unique<ListedForcesGpu>(
1669 fr->ic->epsfac * fr->fudgeQQ,
1670 deviceStreamManager->context(),
1671 deviceStreamManager->bondedStream(havePPDomainDecomposition(cr)),
1673 fr->listedForcesGpu = listedForcesGpu.get();
1676 /* Initialize the mdAtoms structure.
1677 * mdAtoms is not filled with atom data,
1678 * as this can not be done now with domain decomposition.
1680 mdAtoms = makeMDAtoms(fplog, mtop, *inputrec, thisRankHasPmeGpuTask);
1681 if (globalState && thisRankHasPmeGpuTask)
1683 // The pinning of coordinates in the global state object works, because we only use
1684 // PME on GPU without DD or on a separate PME rank, and because the local state pointer
1685 // points to the global state object without DD.
1686 // FIXME: MD and EM separately set up the local state - this should happen in the same
1687 // function, which should also perform the pinning.
1688 changePinningPolicy(&globalState->x, pme_get_pinning_policy());
1691 /* Initialize the virtual site communication */
1692 vsite = makeVirtualSitesHandler(
1693 mtop, cr, fr->pbcType, updateGroups.updateGroupingPerMoleculeType());
1695 calc_shifts(box, fr->shift_vec);
1697 /* With periodic molecules the charge groups should be whole at start up
1698 * and the virtual sites should not be far from their proper positions.
1700 if (!inputrec->bContinuation && MASTER(cr)
1701 && !(inputrec->pbcType != PbcType::No && inputrec->bPeriodicMols))
1703 /* Make molecules whole at start of run */
1704 if (fr->pbcType != PbcType::No)
1706 do_pbc_first_mtop(fplog, inputrec->pbcType, box, &mtop, globalState->x.rvec_array());
1710 /* Correct initial vsite positions are required
1711 * for the initial distribution in the domain decomposition
1712 * and for the initial shell prediction.
1714 constructVirtualSitesGlobal(mtop, globalState->x);
1717 // Make the DD reverse topology, now that any vsites that are present are available
1718 if (DOMAINDECOMP(cr))
1720 dd_make_reverse_top(fplog, cr->dd, mtop, vsite.get(), *inputrec, domdecOptions.ddBondedChecking);
1723 if (EEL_PME(fr->ic->eeltype) || EVDW_PME(fr->ic->vdwtype))
1725 ewaldcoeff_q = fr->ic->ewaldcoeff_q;
1726 ewaldcoeff_lj = fr->ic->ewaldcoeff_lj;
1731 /* This is a PME only node */
1733 GMX_ASSERT(globalState == nullptr,
1734 "We don't need the state on a PME only rank and expect it to be unitialized");
1736 ewaldcoeff_q = calc_ewaldcoeff_q(inputrec->rcoulomb, inputrec->ewald_rtol);
1737 ewaldcoeff_lj = calc_ewaldcoeff_lj(inputrec->rvdw, inputrec->ewald_rtol_lj);
1740 gmx_pme_t* sepPmeData = nullptr;
1741 // This reference hides the fact that PME data is owned by runner on PME-only ranks and by forcerec on other ranks
1742 GMX_ASSERT(thisRankHasDuty(cr, DUTY_PP) == (fr != nullptr),
1743 "Double-checking that only PME-only ranks have no forcerec");
1744 gmx_pme_t*& pmedata = fr ? fr->pmedata : sepPmeData;
1746 // TODO should live in ewald module once its testing is improved
1748 // Later, this program could contain kernels that might be later
1749 // re-used as auto-tuning progresses, or subsequent simulations
1751 PmeGpuProgramStorage pmeGpuProgram;
1752 if (thisRankHasPmeGpuTask)
1755 (deviceStreamManager != nullptr),
1756 "GPU device stream manager should be initialized in order to use GPU for PME.");
1757 GMX_RELEASE_ASSERT((deviceInfo != nullptr),
1758 "GPU device should be initialized in order to use GPU for PME.");
1759 pmeGpuProgram = buildPmeGpuProgram(deviceStreamManager->context());
1762 /* Initiate PME if necessary,
1763 * either on all nodes or on dedicated PME nodes only. */
1764 if (EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype))
1766 if (mdAtoms && mdAtoms->mdatoms())
1768 nChargePerturbed = mdAtoms->mdatoms()->nChargePerturbed;
1769 if (EVDW_PME(inputrec->vdwtype))
1771 nTypePerturbed = mdAtoms->mdatoms()->nTypePerturbed;
1774 if (cr->npmenodes > 0)
1776 /* The PME only nodes need to know nChargePerturbed(FEP on Q) and nTypePerturbed(FEP on LJ)*/
1777 gmx_bcast(sizeof(nChargePerturbed), &nChargePerturbed, cr->mpi_comm_mysim);
1778 gmx_bcast(sizeof(nTypePerturbed), &nTypePerturbed, cr->mpi_comm_mysim);
1781 if (thisRankHasDuty(cr, DUTY_PME))
1785 // TODO: This should be in the builder.
1786 GMX_RELEASE_ASSERT(!runScheduleWork.simulationWork.useGpuPme
1787 || (deviceStreamManager != nullptr),
1788 "Device stream manager should be valid in order to use GPU "
1791 !runScheduleWork.simulationWork.useGpuPme
1792 || deviceStreamManager->streamIsValid(DeviceStreamType::Pme),
1793 "GPU PME stream should be valid in order to use GPU version of PME.");
1795 const DeviceContext* deviceContext = runScheduleWork.simulationWork.useGpuPme
1796 ? &deviceStreamManager->context()
1798 const DeviceStream* pmeStream =
1799 runScheduleWork.simulationWork.useGpuPme
1800 ? &deviceStreamManager->stream(DeviceStreamType::Pme)
1803 pmedata = gmx_pme_init(cr,
1804 getNumPmeDomains(cr->dd),
1806 nChargePerturbed != 0,
1807 nTypePerturbed != 0,
1808 mdrunOptions.reproducible,
1811 gmx_omp_nthreads_get(ModuleMultiThread::Pme),
1816 pmeGpuProgram.get(),
1819 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1824 if (EI_DYNAMICS(inputrec->eI))
1826 /* Turn on signal handling on all nodes */
1828 * (A user signal from the PME nodes (if any)
1829 * is communicated to the PP nodes.
1831 signal_handler_install();
1834 pull_t* pull_work = nullptr;
1835 if (thisRankHasDuty(cr, DUTY_PP))
1837 /* Assumes uniform use of the number of OpenMP threads */
1838 walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(ModuleMultiThread::Default));
1840 if (inputrec->bPull)
1842 /* Initialize pull code */
1843 pull_work = init_pull(fplog,
1844 inputrec->pull.get(),
1849 inputrec->fepvals->init_lambda);
1850 if (inputrec->pull->bXOutAverage || inputrec->pull->bFOutAverage)
1852 initPullHistory(pull_work, &observablesHistory);
1854 if (EI_DYNAMICS(inputrec->eI) && MASTER(cr))
1856 init_pull_output_files(pull_work, filenames.size(), filenames.data(), oenv, startingBehavior);
1860 std::unique_ptr<EnforcedRotation> enforcedRotation;
1863 /* Initialize enforced rotation code */
1864 enforcedRotation = init_rot(fplog,
1877 t_swap* swap = nullptr;
1878 if (inputrec->eSwapCoords != SwapType::No)
1880 /* Initialize ion swapping code */
1881 swap = init_swapcoords(fplog,
1883 opt2fn_master("-swap", filenames.size(), filenames.data(), cr),
1886 &observablesHistory,
1894 /* Let makeConstraints know whether we have essential dynamics constraints. */
1895 auto constr = makeConstraints(mtop,
1898 doEssentialDynamics,
1901 updateGroups.useUpdateGroups(),
1907 /* Energy terms and groups */
1908 gmx_enerdata_t enerd(mtop.groups.groups[SimulationAtomGroupType::EnergyOutput].size(),
1909 inputrec->fepvals->n_lambda);
1911 // cos acceleration is only supported by md, but older tpr
1912 // files might still combine it with other integrators
1913 GMX_RELEASE_ASSERT(inputrec->cos_accel == 0.0 || inputrec->eI == IntegrationAlgorithm::MD,
1914 "cos_acceleration is only supported by integrator=md");
1916 /* Kinetic energy data */
1917 gmx_ekindata_t ekind(inputrec->opts.ngtc,
1918 inputrec->cos_accel,
1919 gmx_omp_nthreads_get(ModuleMultiThread::Update));
1921 /* Set up interactive MD (IMD) */
1922 auto imdSession = makeImdSession(inputrec.get(),
1929 MASTER(cr) ? globalState->x : gmx::ArrayRef<gmx::RVec>(),
1933 mdrunOptions.imdOptions,
1936 if (DOMAINDECOMP(cr))
1938 GMX_RELEASE_ASSERT(fr, "fr was NULL while cr->duty was DUTY_PP");
1939 /* This call is not included in init_domain_decomposition
1940 * because fr->cginfo_mb is set later.
1942 makeBondedLinks(cr->dd, mtop, fr->cginfo_mb);
1945 if (runScheduleWork.simulationWork.useGpuBufferOps)
1947 fr->gpuForceReduction[gmx::AtomLocality::Local] = std::make_unique<gmx::GpuForceReduction>(
1948 deviceStreamManager->context(),
1949 deviceStreamManager->stream(gmx::DeviceStreamType::NonBondedLocal),
1951 fr->gpuForceReduction[gmx::AtomLocality::NonLocal] = std::make_unique<gmx::GpuForceReduction>(
1952 deviceStreamManager->context(),
1953 deviceStreamManager->stream(gmx::DeviceStreamType::NonBondedNonLocal),
1957 std::unique_ptr<gmx::StatePropagatorDataGpu> stateGpu;
1958 if (gpusWereDetected
1959 && ((runScheduleWork.simulationWork.useGpuPme && thisRankHasDuty(cr, DUTY_PME))
1960 || runScheduleWork.simulationWork.useGpuBufferOps))
1962 GpuApiCallBehavior transferKind =
1963 (inputrec->eI == IntegrationAlgorithm::MD && !doRerun && !useModularSimulator)
1964 ? GpuApiCallBehavior::Async
1965 : GpuApiCallBehavior::Sync;
1966 GMX_RELEASE_ASSERT(deviceStreamManager != nullptr,
1967 "GPU device stream manager should be initialized to use GPU.");
1968 stateGpu = std::make_unique<gmx::StatePropagatorDataGpu>(
1969 *deviceStreamManager, transferKind, pme_gpu_get_block_size(fr->pmedata), wcycle.get());
1970 fr->stateGpu = stateGpu.get();
1973 GMX_ASSERT(stopHandlerBuilder_, "Runner must provide StopHandlerBuilder to simulator.");
1974 SimulatorBuilder simulatorBuilder;
1976 simulatorBuilder.add(SimulatorStateData(globalState.get(), &observablesHistory, &enerd, &ekind));
1977 simulatorBuilder.add(std::move(membedHolder));
1978 simulatorBuilder.add(std::move(stopHandlerBuilder_));
1979 simulatorBuilder.add(SimulatorConfig(mdrunOptions, startingBehavior, &runScheduleWork));
1982 simulatorBuilder.add(SimulatorEnv(fplog, cr, ms, mdlog, oenv));
1983 simulatorBuilder.add(Profiling(&nrnb, walltime_accounting, wcycle.get()));
1984 simulatorBuilder.add(ConstraintsParam(
1985 constr.get(), enforcedRotation ? enforcedRotation->getLegacyEnfrot() : nullptr, vsite.get()));
1986 // TODO: Separate `fr` to a separate add, and make the `build` handle the coupling sensibly.
1987 simulatorBuilder.add(LegacyInput(
1988 static_cast<int>(filenames.size()), filenames.data(), inputrec.get(), fr.get()));
1989 simulatorBuilder.add(ReplicaExchangeParameters(replExParams));
1990 simulatorBuilder.add(InteractiveMD(imdSession.get()));
1991 simulatorBuilder.add(SimulatorModules(mdModules_->outputProvider(), mdModules_->notifiers()));
1992 simulatorBuilder.add(CenterOfMassPulling(pull_work));
1993 // Todo move to an MDModule
1994 simulatorBuilder.add(IonSwapping(swap));
1995 simulatorBuilder.add(TopologyData(mtop, mdAtoms.get()));
1996 simulatorBuilder.add(BoxDeformationHandle(deform.get()));
1997 simulatorBuilder.add(std::move(modularSimulatorCheckpointData));
1999 // build and run simulator object based on user-input
2000 auto simulator = simulatorBuilder.build(useModularSimulator);
2003 if (fr->pmePpCommGpu)
2005 // destroy object since it is no longer required. (This needs to be done while the GPU context still exists.)
2006 fr->pmePpCommGpu.reset();
2009 if (inputrec->bPull)
2011 finish_pull(pull_work);
2013 finish_swapcoords(swap);
2017 GMX_RELEASE_ASSERT(pmedata, "pmedata was NULL while cr->duty was not DUTY_PP");
2019 walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(ModuleMultiThread::Pme));
2020 gmx_pmeonly(pmedata,
2024 walltime_accounting,
2027 runScheduleWork.simulationWork.useGpuPmePpCommunication,
2028 deviceStreamManager.get());
2031 wallcycle_stop(wcycle.get(), WallCycleCounter::Run);
2033 /* Finish up, write some stuff
2034 * if rerunMD, don't write last frame again
2042 walltime_accounting,
2043 fr ? fr->nbv.get() : nullptr,
2045 EI_DYNAMICS(inputrec->eI) && !isMultiSim(ms));
2048 deviceStreamManager.reset(nullptr);
2052 gmx_pme_destroy(pmedata);
2056 // FIXME: this is only here to manually unpin mdAtoms->chargeA_ and state->x,
2057 // before we destroy the GPU context(s)
2058 // Pinned buffers are associated with contexts in CUDA.
2059 // As soon as we destroy GPU contexts after mdrunner() exits, these lines should go.
2060 mdAtoms.reset(nullptr);
2061 globalState.reset(nullptr);
2062 mdModules_.reset(nullptr); // destruct force providers here as they might also use the GPU
2063 listedForcesGpu.reset(nullptr);
2064 fr.reset(nullptr); // destruct forcerec before gpu
2065 // TODO convert to C++ so we can get rid of these frees
2068 if (!hwinfo_->deviceInfoList.empty())
2070 /* stop the GPU profiler (only CUDA) */
2074 /* With tMPI we need to wait for all ranks to finish deallocation before
2075 * destroying the CUDA context as some tMPI ranks may be sharing
2078 * This is not a concern in OpenCL where we use one context per rank.
2080 * Note: it is safe to not call the barrier on the ranks which do not use GPU,
2081 * but it is easier and more futureproof to call it on the whole node.
2083 * Note that this function needs to be called even if GPUs are not used
2084 * in this run because the PME ranks have no knowledge of whether GPUs
2085 * are used or not, but all ranks need to enter the barrier below.
2086 * \todo Remove this physical node barrier after making sure
2087 * that it's not needed anymore (with a shared GPU run).
2091 physicalNodeComm.barrier();
2094 if (!devFlags.usingCudaAwareMpi)
2096 // Don't reset GPU in case of CUDA-AWARE MPI
2097 // UCX creates CUDA buffers which are cleaned-up as part of MPI_Finalize()
2098 // resetting the device before MPI_Finalize() results in crashes inside UCX
2099 releaseDevice(deviceInfo);
2102 /* Does what it says */
2103 print_date_and_time(fplog, cr->nodeid, "Finished mdrun", gmx_gettime());
2104 walltime_accounting_destroy(walltime_accounting);
2106 // Ensure log file content is written
2109 gmx_fio_flush(logFileHandle);
2112 /* Reset FPEs (important for unit tests) by disabling them. Assumes no
2113 * exceptions were enabled before function was called. */
2116 gmx_fedisableexcept();
2119 auto rc = static_cast<int>(gmx_get_stop_condition());
2122 /* we need to join all threads. The sub-threads join when they
2123 exit this function, but the master thread needs to be told to
2133 Mdrunner::~Mdrunner()
2135 // Clean up of the Manager.
2136 // This will end up getting called on every thread-MPI rank, which is unnecessary,
2137 // but okay as long as threads synchronize some time before adding or accessing
2138 // a new set of restraints.
2139 if (restraintManager_)
2141 restraintManager_->clear();
2142 GMX_ASSERT(restraintManager_->countRestraints() == 0,
2143 "restraints added during runner life time should be cleared at runner "
2148 void Mdrunner::addPotential(std::shared_ptr<gmx::IRestraintPotential> puller, const std::string& name)
2150 GMX_ASSERT(restraintManager_, "Mdrunner must have a restraint manager.");
2151 // Not sure if this should be logged through the md logger or something else,
2152 // but it is helpful to have some sort of INFO level message sent somewhere.
2153 // std::cout << "Registering restraint named " << name << std::endl;
2155 // When multiple restraints are used, it may be wasteful to register them separately.
2156 // Maybe instead register an entire Restraint Manager as a force provider.
2157 restraintManager_->addToSpec(std::move(puller), name);
2160 Mdrunner::Mdrunner(std::unique_ptr<MDModules> mdModules) : mdModules_(std::move(mdModules)) {}
2162 Mdrunner::Mdrunner(Mdrunner&&) noexcept = default;
2164 //NOLINTNEXTLINE(performance-noexcept-move-constructor) working around GCC bug 58265 in CentOS 7
2165 Mdrunner& Mdrunner::operator=(Mdrunner&& /*handle*/) noexcept(BUGFREE_NOEXCEPT_STRING) = default;
2167 class Mdrunner::BuilderImplementation
2170 BuilderImplementation() = delete;
2171 BuilderImplementation(std::unique_ptr<MDModules> mdModules, compat::not_null<SimulationContext*> context);
2172 ~BuilderImplementation();
2174 BuilderImplementation& setExtraMdrunOptions(const MdrunOptions& options,
2175 real forceWarningThreshold,
2176 StartingBehavior startingBehavior);
2178 void addHardwareDetectionResult(const gmx_hw_info_t* hwinfo);
2180 void addDomdec(const DomdecOptions& options);
2182 void addInput(SimulationInputHandle inputHolder);
2184 void addVerletList(int nstlist);
2186 void addReplicaExchange(const ReplicaExchangeParameters& params);
2188 void addNonBonded(const char* nbpu_opt);
2190 void addPME(const char* pme_opt_, const char* pme_fft_opt_);
2192 void addBondedTaskAssignment(const char* bonded_opt);
2194 void addUpdateTaskAssignment(const char* update_opt);
2196 void addHardwareOptions(const gmx_hw_opt_t& hardwareOptions);
2198 void addFilenames(ArrayRef<const t_filenm> filenames);
2200 void addOutputEnvironment(gmx_output_env_t* outputEnvironment);
2202 void addLogFile(t_fileio* logFileHandle);
2204 void addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder);
2209 // Default parameters copied from runner.h
2210 // \todo Clarify source(s) of default parameters.
2212 const char* nbpu_opt_ = nullptr;
2213 const char* pme_opt_ = nullptr;
2214 const char* pme_fft_opt_ = nullptr;
2215 const char* bonded_opt_ = nullptr;
2216 const char* update_opt_ = nullptr;
2218 MdrunOptions mdrunOptions_;
2220 DomdecOptions domdecOptions_;
2222 ReplicaExchangeParameters replicaExchangeParameters_;
2224 //! Command-line override for the duration of a neighbor list with the Verlet scheme.
2227 //! World communicator, used for hardware detection and task assignment
2228 MPI_Comm libraryWorldCommunicator_ = MPI_COMM_NULL;
2230 //! Multisim communicator handle.
2231 gmx_multisim_t* multiSimulation_;
2233 //! mdrun communicator
2234 MPI_Comm simulationCommunicator_ = MPI_COMM_NULL;
2236 //! Print a warning if any force is larger than this (in kJ/mol nm).
2237 real forceWarningThreshold_ = -1;
2239 //! Whether the simulation will start afresh, or restart with/without appending.
2240 StartingBehavior startingBehavior_ = StartingBehavior::NewSimulation;
2242 //! The modules that comprise the functionality of mdrun.
2243 std::unique_ptr<MDModules> mdModules_;
2245 //! Detected hardware.
2246 const gmx_hw_info_t* hwinfo_ = nullptr;
2248 //! \brief Parallelism information.
2249 gmx_hw_opt_t hardwareOptions_;
2251 //! filename options for simulation.
2252 ArrayRef<const t_filenm> filenames_;
2254 /*! \brief Handle to output environment.
2256 * \todo gmx_output_env_t needs lifetime management.
2258 gmx_output_env_t* outputEnvironment_ = nullptr;
2260 /*! \brief Non-owning handle to MD log file.
2262 * \todo Context should own output facilities for client.
2263 * \todo Improve log file handle management.
2265 * Code managing the FILE* relies on the ability to set it to
2266 * nullptr to check whether the filehandle is valid.
2268 t_fileio* logFileHandle_ = nullptr;
2271 * \brief Builder for simulation stop signal handler.
2273 std::unique_ptr<StopHandlerBuilder> stopHandlerBuilder_ = nullptr;
2276 * \brief Sources for initial simulation state.
2278 * See issue #3652 for near-term refinements to the SimulationInput interface.
2280 * See issue #3379 for broader discussion on API aspects of simulation inputs and outputs.
2282 SimulationInputHandle inputHolder_;
2285 Mdrunner::BuilderImplementation::BuilderImplementation(std::unique_ptr<MDModules> mdModules,
2286 compat::not_null<SimulationContext*> context) :
2287 mdModules_(std::move(mdModules))
2289 libraryWorldCommunicator_ = context->libraryWorldCommunicator_;
2290 simulationCommunicator_ = context->simulationCommunicator_;
2291 multiSimulation_ = context->multiSimulation_.get();
2294 Mdrunner::BuilderImplementation::~BuilderImplementation() = default;
2296 Mdrunner::BuilderImplementation&
2297 Mdrunner::BuilderImplementation::setExtraMdrunOptions(const MdrunOptions& options,
2298 const real forceWarningThreshold,
2299 const StartingBehavior startingBehavior)
2301 mdrunOptions_ = options;
2302 forceWarningThreshold_ = forceWarningThreshold;
2303 startingBehavior_ = startingBehavior;
2307 void Mdrunner::BuilderImplementation::addDomdec(const DomdecOptions& options)
2309 domdecOptions_ = options;
2312 void Mdrunner::BuilderImplementation::addVerletList(int nstlist)
2317 void Mdrunner::BuilderImplementation::addReplicaExchange(const ReplicaExchangeParameters& params)
2319 replicaExchangeParameters_ = params;
2322 Mdrunner Mdrunner::BuilderImplementation::build()
2324 auto newRunner = Mdrunner(std::move(mdModules_));
2326 newRunner.mdrunOptions = mdrunOptions_;
2327 newRunner.pforce = forceWarningThreshold_;
2328 newRunner.startingBehavior = startingBehavior_;
2329 newRunner.domdecOptions = domdecOptions_;
2331 // \todo determine an invariant to check or confirm that all gmx_hw_opt_t objects are valid
2332 newRunner.hw_opt = hardwareOptions_;
2334 // No invariant to check. This parameter exists to optionally override other behavior.
2335 newRunner.nstlist_cmdline = nstlist_;
2337 newRunner.replExParams = replicaExchangeParameters_;
2339 newRunner.filenames = filenames_;
2341 newRunner.libraryWorldCommunicator = libraryWorldCommunicator_;
2343 newRunner.simulationCommunicator = simulationCommunicator_;
2345 // nullptr is a valid value for the multisim handle
2346 newRunner.ms = multiSimulation_;
2350 newRunner.hwinfo_ = hwinfo_;
2354 GMX_THROW(gmx::APIError(
2355 "MdrunnerBuilder::addHardwareDetectionResult() is required before build()"));
2360 newRunner.inputHolder_ = std::move(inputHolder_);
2364 GMX_THROW(gmx::APIError("MdrunnerBuilder::addInput() is required before build()."));
2367 // \todo Clarify ownership and lifetime management for gmx_output_env_t
2368 // \todo Update sanity checking when output environment has clearly specified invariants.
2369 // Initialization and default values for oenv are not well specified in the current version.
2370 if (outputEnvironment_)
2372 newRunner.oenv = outputEnvironment_;
2376 GMX_THROW(gmx::APIError(
2377 "MdrunnerBuilder::addOutputEnvironment() is required before build()"));
2380 newRunner.logFileHandle = logFileHandle_;
2384 newRunner.nbpu_opt = nbpu_opt_;
2388 GMX_THROW(gmx::APIError("MdrunnerBuilder::addNonBonded() is required before build()"));
2391 if (pme_opt_ && pme_fft_opt_)
2393 newRunner.pme_opt = pme_opt_;
2394 newRunner.pme_fft_opt = pme_fft_opt_;
2398 GMX_THROW(gmx::APIError("MdrunnerBuilder::addElectrostatics() is required before build()"));
2403 newRunner.bonded_opt = bonded_opt_;
2407 GMX_THROW(gmx::APIError(
2408 "MdrunnerBuilder::addBondedTaskAssignment() is required before build()"));
2413 newRunner.update_opt = update_opt_;
2417 GMX_THROW(gmx::APIError(
2418 "MdrunnerBuilder::addUpdateTaskAssignment() is required before build() "));
2422 newRunner.restraintManager_ = std::make_unique<gmx::RestraintManager>();
2424 if (stopHandlerBuilder_)
2426 newRunner.stopHandlerBuilder_ = std::move(stopHandlerBuilder_);
2430 newRunner.stopHandlerBuilder_ = std::make_unique<StopHandlerBuilder>();
2436 void Mdrunner::BuilderImplementation::addHardwareDetectionResult(const gmx_hw_info_t* hwinfo)
2441 void Mdrunner::BuilderImplementation::addNonBonded(const char* nbpu_opt)
2443 nbpu_opt_ = nbpu_opt;
2446 void Mdrunner::BuilderImplementation::addPME(const char* pme_opt, const char* pme_fft_opt)
2449 pme_fft_opt_ = pme_fft_opt;
2452 void Mdrunner::BuilderImplementation::addBondedTaskAssignment(const char* bonded_opt)
2454 bonded_opt_ = bonded_opt;
2457 void Mdrunner::BuilderImplementation::addUpdateTaskAssignment(const char* update_opt)
2459 update_opt_ = update_opt;
2462 void Mdrunner::BuilderImplementation::addHardwareOptions(const gmx_hw_opt_t& hardwareOptions)
2464 hardwareOptions_ = hardwareOptions;
2467 void Mdrunner::BuilderImplementation::addFilenames(ArrayRef<const t_filenm> filenames)
2469 filenames_ = filenames;
2472 void Mdrunner::BuilderImplementation::addOutputEnvironment(gmx_output_env_t* outputEnvironment)
2474 outputEnvironment_ = outputEnvironment;
2477 void Mdrunner::BuilderImplementation::addLogFile(t_fileio* logFileHandle)
2479 logFileHandle_ = logFileHandle;
2482 void Mdrunner::BuilderImplementation::addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder)
2484 stopHandlerBuilder_ = std::move(builder);
2487 void Mdrunner::BuilderImplementation::addInput(SimulationInputHandle inputHolder)
2489 inputHolder_ = std::move(inputHolder);
2492 MdrunnerBuilder::MdrunnerBuilder(std::unique_ptr<MDModules> mdModules,
2493 compat::not_null<SimulationContext*> context) :
2494 impl_{ std::make_unique<Mdrunner::BuilderImplementation>(std::move(mdModules), context) }
2498 MdrunnerBuilder::~MdrunnerBuilder() = default;
2500 MdrunnerBuilder& MdrunnerBuilder::addHardwareDetectionResult(const gmx_hw_info_t* hwinfo)
2502 impl_->addHardwareDetectionResult(hwinfo);
2506 MdrunnerBuilder& MdrunnerBuilder::addSimulationMethod(const MdrunOptions& options,
2507 real forceWarningThreshold,
2508 const StartingBehavior startingBehavior)
2510 impl_->setExtraMdrunOptions(options, forceWarningThreshold, startingBehavior);
2514 MdrunnerBuilder& MdrunnerBuilder::addDomainDecomposition(const DomdecOptions& options)
2516 impl_->addDomdec(options);
2520 MdrunnerBuilder& MdrunnerBuilder::addNeighborList(int nstlist)
2522 impl_->addVerletList(nstlist);
2526 MdrunnerBuilder& MdrunnerBuilder::addReplicaExchange(const ReplicaExchangeParameters& params)
2528 impl_->addReplicaExchange(params);
2532 MdrunnerBuilder& MdrunnerBuilder::addNonBonded(const char* nbpu_opt)
2534 impl_->addNonBonded(nbpu_opt);
2538 MdrunnerBuilder& MdrunnerBuilder::addElectrostatics(const char* pme_opt, const char* pme_fft_opt)
2540 // The builder method may become more general in the future, but in this version,
2541 // parameters for PME electrostatics are both required and the only parameters
2543 if (pme_opt && pme_fft_opt)
2545 impl_->addPME(pme_opt, pme_fft_opt);
2550 gmx::InvalidInputError("addElectrostatics() arguments must be non-null pointers."));
2555 MdrunnerBuilder& MdrunnerBuilder::addBondedTaskAssignment(const char* bonded_opt)
2557 impl_->addBondedTaskAssignment(bonded_opt);
2561 MdrunnerBuilder& MdrunnerBuilder::addUpdateTaskAssignment(const char* update_opt)
2563 impl_->addUpdateTaskAssignment(update_opt);
2567 Mdrunner MdrunnerBuilder::build()
2569 return impl_->build();
2572 MdrunnerBuilder& MdrunnerBuilder::addHardwareOptions(const gmx_hw_opt_t& hardwareOptions)
2574 impl_->addHardwareOptions(hardwareOptions);
2578 MdrunnerBuilder& MdrunnerBuilder::addFilenames(ArrayRef<const t_filenm> filenames)
2580 impl_->addFilenames(filenames);
2584 MdrunnerBuilder& MdrunnerBuilder::addOutputEnvironment(gmx_output_env_t* outputEnvironment)
2586 impl_->addOutputEnvironment(outputEnvironment);
2590 MdrunnerBuilder& MdrunnerBuilder::addLogFile(t_fileio* logFileHandle)
2592 impl_->addLogFile(logFileHandle);
2596 MdrunnerBuilder& MdrunnerBuilder::addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder)
2598 impl_->addStopHandlerBuilder(std::move(builder));
2602 MdrunnerBuilder& MdrunnerBuilder::addInput(SimulationInputHandle input)
2604 impl_->addInput(std::move(input));
2608 MdrunnerBuilder::MdrunnerBuilder(MdrunnerBuilder&&) noexcept = default;
2610 MdrunnerBuilder& MdrunnerBuilder::operator=(MdrunnerBuilder&&) noexcept = default;