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/makebondedlinks.h"
66 #include "gromacs/domdec/partition.h"
67 #include "gromacs/domdec/reversetopology.h"
68 #include "gromacs/ewald/ewald_utils.h"
69 #include "gromacs/ewald/pme.h"
70 #include "gromacs/ewald/pme_gpu_program.h"
71 #include "gromacs/ewald/pme_only.h"
72 #include "gromacs/ewald/pme_pp_comm_gpu.h"
73 #include "gromacs/fileio/checkpoint.h"
74 #include "gromacs/fileio/gmxfio.h"
75 #include "gromacs/fileio/oenv.h"
76 #include "gromacs/fileio/tpxio.h"
77 #include "gromacs/gmxlib/network.h"
78 #include "gromacs/gmxlib/nrnb.h"
79 #include "gromacs/gpu_utils/device_stream_manager.h"
80 #include "gromacs/hardware/cpuinfo.h"
81 #include "gromacs/hardware/detecthardware.h"
82 #include "gromacs/hardware/device_management.h"
83 #include "gromacs/hardware/hardwaretopology.h"
84 #include "gromacs/hardware/printhardware.h"
85 #include "gromacs/imd/imd.h"
86 #include "gromacs/listed_forces/disre.h"
87 #include "gromacs/listed_forces/listed_forces_gpu.h"
88 #include "gromacs/listed_forces/listed_forces.h"
89 #include "gromacs/listed_forces/orires.h"
90 #include "gromacs/math/functions.h"
91 #include "gromacs/math/utilities.h"
92 #include "gromacs/math/vec.h"
93 #include "gromacs/mdlib/boxdeformation.h"
94 #include "gromacs/mdlib/broadcaststructs.h"
95 #include "gromacs/mdlib/calc_verletbuf.h"
96 #include "gromacs/mdlib/dispersioncorrection.h"
97 #include "gromacs/mdlib/enerdata_utils.h"
98 #include "gromacs/mdlib/force.h"
99 #include "gromacs/mdlib/forcerec.h"
100 #include "gromacs/mdlib/gmx_omp_nthreads.h"
101 #include "gromacs/mdlib/gpuforcereduction.h"
102 #include "gromacs/mdlib/makeconstraints.h"
103 #include "gromacs/mdlib/md_support.h"
104 #include "gromacs/mdlib/mdatoms.h"
105 #include "gromacs/mdlib/sighandler.h"
106 #include "gromacs/mdlib/stophandler.h"
107 #include "gromacs/mdlib/tgroup.h"
108 #include "gromacs/mdlib/updategroups.h"
109 #include "gromacs/mdlib/vsite.h"
110 #include "gromacs/mdrun/mdmodules.h"
111 #include "gromacs/mdrun/simulationcontext.h"
112 #include "gromacs/mdrun/simulationinput.h"
113 #include "gromacs/mdrun/simulationinputhandle.h"
114 #include "gromacs/mdrunutility/handlerestart.h"
115 #include "gromacs/mdrunutility/logging.h"
116 #include "gromacs/mdrunutility/multisim.h"
117 #include "gromacs/mdrunutility/printtime.h"
118 #include "gromacs/mdrunutility/threadaffinity.h"
119 #include "gromacs/mdtypes/checkpointdata.h"
120 #include "gromacs/mdtypes/commrec.h"
121 #include "gromacs/mdtypes/enerdata.h"
122 #include "gromacs/mdtypes/fcdata.h"
123 #include "gromacs/mdtypes/forcerec.h"
124 #include "gromacs/mdtypes/group.h"
125 #include "gromacs/mdtypes/inputrec.h"
126 #include "gromacs/mdtypes/interaction_const.h"
127 #include "gromacs/mdtypes/md_enums.h"
128 #include "gromacs/mdtypes/mdatom.h"
129 #include "gromacs/mdtypes/mdrunoptions.h"
130 #include "gromacs/mdtypes/observableshistory.h"
131 #include "gromacs/mdtypes/observablesreducer.h"
132 #include "gromacs/mdtypes/simulation_workload.h"
133 #include "gromacs/mdtypes/state.h"
134 #include "gromacs/mdtypes/state_propagator_data_gpu.h"
135 #include "gromacs/modularsimulator/modularsimulator.h"
136 #include "gromacs/nbnxm/gpu_data_mgmt.h"
137 #include "gromacs/nbnxm/nbnxm.h"
138 #include "gromacs/nbnxm/pairlist_tuning.h"
139 #include "gromacs/pbcutil/pbc.h"
140 #include "gromacs/pulling/output.h"
141 #include "gromacs/pulling/pull.h"
142 #include "gromacs/pulling/pull_rotation.h"
143 #include "gromacs/restraint/manager.h"
144 #include "gromacs/restraint/restraintmdmodule.h"
145 #include "gromacs/restraint/restraintpotential.h"
146 #include "gromacs/swap/swapcoords.h"
147 #include "gromacs/taskassignment/decidegpuusage.h"
148 #include "gromacs/taskassignment/decidesimulationworkload.h"
149 #include "gromacs/taskassignment/resourcedivision.h"
150 #include "gromacs/taskassignment/taskassignment.h"
151 #include "gromacs/taskassignment/usergpuids.h"
152 #include "gromacs/timing/gpu_timing.h"
153 #include "gromacs/timing/wallcycle.h"
154 #include "gromacs/timing/wallcyclereporting.h"
155 #include "gromacs/topology/mtop_util.h"
156 #include "gromacs/trajectory/trajectoryframe.h"
157 #include "gromacs/utility/basenetwork.h"
158 #include "gromacs/utility/cstringutil.h"
159 #include "gromacs/utility/exceptions.h"
160 #include "gromacs/utility/fatalerror.h"
161 #include "gromacs/utility/filestream.h"
162 #include "gromacs/utility/gmxassert.h"
163 #include "gromacs/utility/gmxmpi.h"
164 #include "gromacs/utility/keyvaluetree.h"
165 #include "gromacs/utility/logger.h"
166 #include "gromacs/utility/loggerbuilder.h"
167 #include "gromacs/utility/mdmodulesnotifiers.h"
168 #include "gromacs/utility/physicalnodecommunicator.h"
169 #include "gromacs/utility/pleasecite.h"
170 #include "gromacs/utility/programcontext.h"
171 #include "gromacs/utility/smalloc.h"
172 #include "gromacs/utility/stringutil.h"
173 #include "gromacs/utility/mpiinfo.h"
175 #include "isimulator.h"
176 #include "membedholder.h"
177 #include "replicaexchange.h"
178 #include "simulatorbuilder.h"
184 /*! \brief Manage any development feature flag variables encountered
186 * The use of dev features indicated by environment variables is
187 * logged in order to ensure that runs with such features enabled can
188 * be identified from their log and standard output. Any cross
189 * dependencies are also checked, and if unsatisfied, a fatal error
192 * Note that some development features overrides are applied already here:
193 * the GPU communication flags are set to false in non-tMPI and non-CUDA builds.
195 * \param[in] mdlog Logger object.
196 * \param[in] useGpuForNonbonded True if the nonbonded task is offloaded in this run.
197 * \param[in] pmeRunMode The PME run mode for this run
198 * \returns The object populated with development feature flags.
200 static DevelopmentFeatureFlags manageDevelopmentFeatures(const gmx::MDLogger& mdlog,
201 const bool useGpuForNonbonded,
202 const PmeRunMode pmeRunMode)
204 DevelopmentFeatureFlags devFlags;
206 devFlags.enableGpuBufferOps =
207 GMX_GPU_CUDA && useGpuForNonbonded && (getenv("GMX_USE_GPU_BUFFER_OPS") != nullptr);
208 devFlags.enableGpuHaloExchange = GMX_MPI && GMX_GPU_CUDA && getenv("GMX_GPU_DD_COMMS") != nullptr;
209 devFlags.forceGpuUpdateDefault = (getenv("GMX_FORCE_UPDATE_DEFAULT_GPU") != nullptr) || GMX_FAHCORE;
210 devFlags.enableGpuPmePPComm = GMX_MPI && GMX_GPU_CUDA && getenv("GMX_GPU_PME_PP_COMMS") != nullptr;
212 // Direct GPU comm path is being used with CUDA_AWARE_MPI
213 // make sure underlying MPI implementation is CUDA-aware
214 if (!GMX_THREAD_MPI && (devFlags.enableGpuPmePPComm || devFlags.enableGpuHaloExchange))
216 const bool haveDetectedCudaAwareMpi =
217 (checkMpiCudaAwareSupport() == CudaAwareMpiStatus::Supported);
218 const bool forceCudaAwareMpi = (getenv("GMX_FORCE_CUDA_AWARE_MPI") != nullptr);
220 if (!haveDetectedCudaAwareMpi && forceCudaAwareMpi)
222 // CUDA-aware support not detected in MPI library but, user has forced it's use
223 GMX_LOG(mdlog.warning)
225 .appendTextFormatted(
226 "This run has forced use of 'CUDA-aware MPI'. "
227 "But, GROMACS cannot determine if underlying MPI "
228 "is CUDA-aware. GROMACS recommends use of latest openMPI version "
229 "for CUDA-aware support. "
230 "If you observe failures at runtime, try unsetting "
231 "GMX_FORCE_CUDA_AWARE_MPI environment variable.");
234 if (haveDetectedCudaAwareMpi || forceCudaAwareMpi)
236 devFlags.usingCudaAwareMpi = true;
237 GMX_LOG(mdlog.warning)
239 .appendTextFormatted(
240 "Using CUDA-aware MPI for 'GPU halo exchange' or 'GPU PME-PP "
241 "communications' feature.");
245 if (devFlags.enableGpuHaloExchange)
247 GMX_LOG(mdlog.warning)
249 .appendTextFormatted(
250 "GMX_GPU_DD_COMMS environment variable detected, but the 'GPU "
251 "halo exchange' feature will not be enabled as GROMACS couldn't "
252 "detect CUDA_aware support in underlying MPI implementation.");
253 devFlags.enableGpuHaloExchange = false;
255 if (devFlags.enableGpuPmePPComm)
257 GMX_LOG(mdlog.warning)
260 "GMX_GPU_PME_PP_COMMS environment variable detected, but the "
261 "'GPU PME-PP communications' feature will not be enabled as "
263 "detect CUDA_aware support in underlying MPI implementation.");
264 devFlags.enableGpuPmePPComm = false;
267 GMX_LOG(mdlog.warning)
269 .appendTextFormatted(
270 "GROMACS recommends use of latest OpenMPI version for CUDA-aware "
272 "If you are certain about CUDA-aware support in your MPI library, "
273 "you can force it's use by setting environment variable "
274 " GMX_FORCE_CUDA_AWARE_MPI.");
278 if (devFlags.enableGpuBufferOps)
280 GMX_LOG(mdlog.warning)
282 .appendTextFormatted(
283 "This run uses the 'GPU buffer ops' feature, enabled by the "
284 "GMX_USE_GPU_BUFFER_OPS environment variable.");
287 if (devFlags.forceGpuUpdateDefault)
289 GMX_LOG(mdlog.warning)
291 .appendTextFormatted(
292 "This run will default to '-update gpu' as requested by the "
293 "GMX_FORCE_UPDATE_DEFAULT_GPU environment variable. GPU update with domain "
294 "decomposition lacks substantial testing and should be used with caution.");
297 if (devFlags.enableGpuHaloExchange)
299 if (useGpuForNonbonded)
301 if (!devFlags.enableGpuBufferOps)
303 GMX_LOG(mdlog.warning)
305 .appendTextFormatted(
306 "Enabling GPU buffer operations required by GMX_GPU_DD_COMMS "
307 "(equivalent with GMX_USE_GPU_BUFFER_OPS=1).");
308 devFlags.enableGpuBufferOps = true;
310 GMX_LOG(mdlog.warning)
312 .appendTextFormatted(
313 "This run has requested the 'GPU halo exchange' feature, enabled by "
315 "GMX_GPU_DD_COMMS environment variable.");
319 GMX_LOG(mdlog.warning)
321 .appendTextFormatted(
322 "GMX_GPU_DD_COMMS environment variable detected, but the 'GPU "
323 "halo exchange' feature will not be enabled as nonbonded interactions "
324 "are not offloaded.");
325 devFlags.enableGpuHaloExchange = false;
329 if (devFlags.enableGpuPmePPComm)
331 if (pmeRunMode == PmeRunMode::GPU)
333 if (!devFlags.enableGpuBufferOps)
335 GMX_LOG(mdlog.warning)
337 .appendTextFormatted(
338 "Enabling GPU buffer operations required by GMX_GPU_PME_PP_COMMS "
339 "(equivalent with GMX_USE_GPU_BUFFER_OPS=1).");
340 devFlags.enableGpuBufferOps = true;
342 GMX_LOG(mdlog.warning)
344 .appendTextFormatted(
345 "This run uses the 'GPU PME-PP communications' feature, enabled "
346 "by the GMX_GPU_PME_PP_COMMS environment variable.");
350 std::string clarification;
351 if (pmeRunMode == PmeRunMode::Mixed)
354 "PME FFT and gather are not offloaded to the GPU (PME is running in mixed "
359 clarification = "PME is not offloaded to the GPU.";
361 GMX_LOG(mdlog.warning)
364 "GMX_GPU_PME_PP_COMMS environment variable detected, but the "
365 "'GPU PME-PP communications' feature was not enabled as "
367 devFlags.enableGpuPmePPComm = false;
374 /*! \brief Barrier for safe simultaneous thread access to mdrunner data
376 * Used to ensure that the master thread does not modify mdrunner during copy
377 * on the spawned threads. */
378 static void threadMpiMdrunnerAccessBarrier()
381 MPI_Barrier(MPI_COMM_WORLD);
385 Mdrunner Mdrunner::cloneOnSpawnedThread() const
387 auto newRunner = Mdrunner(std::make_unique<MDModules>());
389 // All runners in the same process share a restraint manager resource because it is
390 // part of the interface to the client code, which is associated only with the
391 // original thread. Handles to the same resources can be obtained by copy.
393 newRunner.restraintManager_ = std::make_unique<RestraintManager>(*restraintManager_);
396 // Copy members of master runner.
397 // \todo Replace with builder when Simulation context and/or runner phases are better defined.
398 // Ref https://gitlab.com/gromacs/gromacs/-/issues/2587 and https://gitlab.com/gromacs/gromacs/-/issues/2375
399 newRunner.hw_opt = hw_opt;
400 newRunner.filenames = filenames;
402 newRunner.hwinfo_ = hwinfo_;
403 newRunner.oenv = oenv;
404 newRunner.mdrunOptions = mdrunOptions;
405 newRunner.domdecOptions = domdecOptions;
406 newRunner.nbpu_opt = nbpu_opt;
407 newRunner.pme_opt = pme_opt;
408 newRunner.pme_fft_opt = pme_fft_opt;
409 newRunner.bonded_opt = bonded_opt;
410 newRunner.update_opt = update_opt;
411 newRunner.nstlist_cmdline = nstlist_cmdline;
412 newRunner.replExParams = replExParams;
413 newRunner.pforce = pforce;
414 // Give the spawned thread the newly created valid communicator
415 // for the simulation.
416 newRunner.libraryWorldCommunicator = MPI_COMM_WORLD;
417 newRunner.simulationCommunicator = MPI_COMM_WORLD;
419 newRunner.startingBehavior = startingBehavior;
420 newRunner.stopHandlerBuilder_ = std::make_unique<StopHandlerBuilder>(*stopHandlerBuilder_);
421 newRunner.inputHolder_ = inputHolder_;
423 threadMpiMdrunnerAccessBarrier();
428 /*! \brief The callback used for running on spawned threads.
430 * Obtains the pointer to the master mdrunner object from the one
431 * argument permitted to the thread-launch API call, copies it to make
432 * a new runner for this thread, reinitializes necessary data, and
433 * proceeds to the simulation. */
434 static void mdrunner_start_fn(const void* arg)
438 const auto* masterMdrunner = reinterpret_cast<const gmx::Mdrunner*>(arg);
439 /* copy the arg list to make sure that it's thread-local. This
440 doesn't copy pointed-to items, of course; fnm, cr and fplog
441 are reset in the call below, all others should be const. */
442 gmx::Mdrunner mdrunner = masterMdrunner->cloneOnSpawnedThread();
445 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
449 void Mdrunner::spawnThreads(int numThreadsToLaunch)
452 /* now spawn new threads that start mdrunner_start_fn(), while
453 the main thread returns. Thread affinity is handled later. */
454 if (tMPI_Init_fn(TRUE, numThreadsToLaunch, TMPI_AFFINITY_NONE, mdrunner_start_fn, static_cast<const void*>(this))
457 GMX_THROW(gmx::InternalError("Failed to spawn thread-MPI threads"));
460 // Give the master thread the newly created valid communicator for
462 libraryWorldCommunicator = MPI_COMM_WORLD;
463 simulationCommunicator = MPI_COMM_WORLD;
464 threadMpiMdrunnerAccessBarrier();
466 GMX_UNUSED_VALUE(numThreadsToLaunch);
467 GMX_UNUSED_VALUE(mdrunner_start_fn);
473 /*! \brief Initialize variables for Verlet scheme simulation */
474 static void prepare_verlet_scheme(FILE* fplog,
478 const gmx_mtop_t& mtop,
480 bool makeGpuPairList,
481 const gmx::CpuInfo& cpuinfo)
483 // We checked the cut-offs in grompp, but double-check here.
484 // We have PME+LJcutoff kernels for rcoulomb>rvdw.
485 if (EEL_PME_EWALD(ir->coulombtype) && ir->vdwtype == VanDerWaalsType::Cut)
487 GMX_RELEASE_ASSERT(ir->rcoulomb >= ir->rvdw,
488 "With Verlet lists and PME we should have rcoulomb>=rvdw");
492 GMX_RELEASE_ASSERT(ir->rcoulomb == ir->rvdw,
493 "With Verlet lists and no PME rcoulomb and rvdw should be identical");
495 /* For NVE simulations, we will retain the initial list buffer */
496 if (EI_DYNAMICS(ir->eI) && ir->verletbuf_tol > 0
497 && !(EI_MD(ir->eI) && ir->etc == TemperatureCoupling::No))
499 /* Update the Verlet buffer size for the current run setup */
501 /* Here we assume SIMD-enabled kernels are being used. But as currently
502 * calc_verlet_buffer_size gives the same results for 4x8 and 4x4
503 * and 4x2 gives a larger buffer than 4x4, this is ok.
505 ListSetupType listType =
506 (makeGpuPairList ? ListSetupType::Gpu : ListSetupType::CpuSimdWhenSupported);
507 VerletbufListSetup listSetup = verletbufGetSafeListSetup(listType);
509 const real rlist_new =
510 calcVerletBufferSize(mtop, det(box), *ir, ir->nstlist, ir->nstlist - 1, -1, listSetup);
512 if (rlist_new != ir->rlist)
514 if (fplog != nullptr)
517 "\nChanging rlist from %g to %g for non-bonded %dx%d atom kernels\n\n",
520 listSetup.cluster_size_i,
521 listSetup.cluster_size_j);
523 ir->rlist = rlist_new;
527 if (nstlist_cmdline > 0 && (!EI_DYNAMICS(ir->eI) || ir->verletbuf_tol <= 0))
530 "Can not set nstlist without %s",
531 !EI_DYNAMICS(ir->eI) ? "dynamics" : "verlet-buffer-tolerance");
534 if (EI_DYNAMICS(ir->eI))
536 /* Set or try nstlist values */
537 increaseNstlist(fplog, cr, ir, nstlist_cmdline, &mtop, box, makeGpuPairList, cpuinfo);
541 /*! \brief Override the nslist value in inputrec
543 * with value passed on the command line (if any)
545 static void override_nsteps_cmdline(const gmx::MDLogger& mdlog, int64_t nsteps_cmdline, t_inputrec* ir)
549 /* override with anything else than the default -2 */
550 if (nsteps_cmdline > -2)
552 char sbuf_steps[STEPSTRSIZE];
553 char sbuf_msg[STRLEN];
555 ir->nsteps = nsteps_cmdline;
556 if (EI_DYNAMICS(ir->eI) && nsteps_cmdline != -1)
559 "Overriding nsteps with value passed on the command line: %s steps, %.3g ps",
560 gmx_step_str(nsteps_cmdline, sbuf_steps),
561 fabs(nsteps_cmdline * ir->delta_t));
566 "Overriding nsteps with value passed on the command line: %s steps",
567 gmx_step_str(nsteps_cmdline, sbuf_steps));
570 GMX_LOG(mdlog.warning).asParagraph().appendText(sbuf_msg);
572 else if (nsteps_cmdline < -2)
574 gmx_fatal(FARGS, "Invalid nsteps value passed on the command line: %" PRId64, nsteps_cmdline);
576 /* Do nothing if nsteps_cmdline == -2 */
582 /*! \brief Return whether GPU acceleration of nonbondeds is supported with the given settings.
584 * If not, and if a warning may be issued, logs a warning about
585 * falling back to CPU code. With thread-MPI, only the first
586 * call to this function should have \c issueWarning true. */
587 static bool gpuAccelerationOfNonbondedIsUseful(const MDLogger& mdlog, const t_inputrec& ir, bool issueWarning)
589 bool gpuIsUseful = true;
592 if (ir.opts.ngener - ir.nwall > 1)
594 /* The GPU code does not support more than one energy group.
595 * If the user requested GPUs explicitly, a fatal error is given later.
599 "Multiple energy groups is not implemented for GPUs, falling back to the CPU. "
600 "For better performance, run on the GPU without energy groups and then do "
601 "gmx mdrun -rerun option on the trajectory with an energy group .tpr file.";
607 warning = "TPI is not implemented for GPUs.";
610 if (!gpuIsUseful && issueWarning)
612 GMX_LOG(mdlog.warning).asParagraph().appendText(warning);
618 //! Initializes the logger for mdrun.
619 static gmx::LoggerOwner buildLogger(FILE* fplog, const bool isSimulationMasterRank)
621 gmx::LoggerBuilder builder;
622 if (fplog != nullptr)
624 builder.addTargetFile(gmx::MDLogger::LogLevel::Info, fplog);
626 if (isSimulationMasterRank)
628 builder.addTargetStream(gmx::MDLogger::LogLevel::Warning, &gmx::TextOutputFile::standardError());
630 return builder.build();
633 //! Make a TaskTarget from an mdrun argument string.
634 static TaskTarget findTaskTarget(const char* optionString)
636 TaskTarget returnValue = TaskTarget::Auto;
638 if (strncmp(optionString, "auto", 3) == 0)
640 returnValue = TaskTarget::Auto;
642 else if (strncmp(optionString, "cpu", 3) == 0)
644 returnValue = TaskTarget::Cpu;
646 else if (strncmp(optionString, "gpu", 3) == 0)
648 returnValue = TaskTarget::Gpu;
652 GMX_ASSERT(false, "Option string should have been checked for sanity already");
658 //! Finish run, aggregate data to print performance info.
659 static void finish_run(FILE* fplog,
660 const gmx::MDLogger& mdlog,
662 const t_inputrec& inputrec,
664 gmx_wallcycle* wcycle,
665 gmx_walltime_accounting_t walltime_accounting,
666 nonbonded_verlet_t* nbv,
667 const gmx_pme_t* pme,
671 double nbfs = 0, mflop = 0;
672 double elapsed_time, elapsed_time_over_all_ranks, elapsed_time_over_all_threads,
673 elapsed_time_over_all_threads_over_all_ranks;
674 /* Control whether it is valid to print a report. Only the
675 simulation master may print, but it should not do so if the run
676 terminated e.g. before a scheduled reset step. This is
677 complicated by the fact that PME ranks are unaware of the
678 reason why they were sent a pmerecvqxFINISH. To avoid
679 communication deadlocks, we always do the communication for the
680 report, even if we've decided not to write the report, because
681 how long it takes to finish the run is not important when we've
682 decided not to report on the simulation performance.
684 Further, we only report performance for dynamical integrators,
685 because those are the only ones for which we plan to
686 consider doing any optimizations. */
687 bool printReport = EI_DYNAMICS(inputrec.eI) && SIMMASTER(cr);
689 if (printReport && !walltime_accounting_get_valid_finish(walltime_accounting))
691 GMX_LOG(mdlog.warning)
693 .appendText("Simulation ended prematurely, no performance report will be written.");
698 std::unique_ptr<t_nrnb> nrnbTotalStorage;
701 nrnbTotalStorage = std::make_unique<t_nrnb>();
702 nrnb_tot = nrnbTotalStorage.get();
704 MPI_Allreduce(nrnb->n.data(), nrnb_tot->n.data(), eNRNB, MPI_DOUBLE, MPI_SUM, cr->mpi_comm_mysim);
712 elapsed_time = walltime_accounting_get_time_since_reset(walltime_accounting);
713 elapsed_time_over_all_threads =
714 walltime_accounting_get_time_since_reset_over_all_threads(walltime_accounting);
718 /* reduce elapsed_time over all MPI ranks in the current simulation */
719 MPI_Allreduce(&elapsed_time, &elapsed_time_over_all_ranks, 1, MPI_DOUBLE, MPI_SUM, cr->mpi_comm_mysim);
720 elapsed_time_over_all_ranks /= cr->nnodes;
721 /* Reduce elapsed_time_over_all_threads over all MPI ranks in the
722 * current simulation. */
723 MPI_Allreduce(&elapsed_time_over_all_threads,
724 &elapsed_time_over_all_threads_over_all_ranks,
733 elapsed_time_over_all_ranks = elapsed_time;
734 elapsed_time_over_all_threads_over_all_ranks = elapsed_time_over_all_threads;
739 print_flop(fplog, nrnb_tot, &nbfs, &mflop);
742 if (thisRankHasDuty(cr, DUTY_PP) && haveDDAtomOrdering(*cr))
744 print_dd_statistics(cr, inputrec, fplog);
747 /* TODO Move the responsibility for any scaling by thread counts
748 * to the code that handled the thread region, so that there's a
749 * mechanism to keep cycle counting working during the transition
750 * to task parallelism. */
751 int nthreads_pp = gmx_omp_nthreads_get(ModuleMultiThread::Nonbonded);
752 int nthreads_pme = gmx_omp_nthreads_get(ModuleMultiThread::Pme);
753 wallcycle_scale_by_num_threads(
754 wcycle, thisRankHasDuty(cr, DUTY_PME) && !thisRankHasDuty(cr, DUTY_PP), nthreads_pp, nthreads_pme);
755 auto cycle_sum(wallcycle_sum(cr, wcycle));
759 auto* nbnxn_gpu_timings =
760 (nbv != nullptr && nbv->useGpu()) ? Nbnxm::gpu_get_timings(nbv->gpu_nbv) : nullptr;
761 gmx_wallclock_gpu_pme_t pme_gpu_timings = {};
763 if (pme_gpu_task_enabled(pme))
765 pme_gpu_get_timings(pme, &pme_gpu_timings);
767 wallcycle_print(fplog,
773 elapsed_time_over_all_ranks,
779 if (EI_DYNAMICS(inputrec.eI))
781 delta_t = inputrec.delta_t;
787 elapsed_time_over_all_threads_over_all_ranks,
788 elapsed_time_over_all_ranks,
789 walltime_accounting_get_nsteps_done_since_reset(walltime_accounting),
797 elapsed_time_over_all_threads_over_all_ranks,
798 elapsed_time_over_all_ranks,
799 walltime_accounting_get_nsteps_done_since_reset(walltime_accounting),
807 int Mdrunner::mdrunner()
810 std::unique_ptr<t_forcerec> fr;
811 real ewaldcoeff_q = 0;
812 real ewaldcoeff_lj = 0;
813 int nChargePerturbed = -1, nTypePerturbed = 0;
814 gmx_walltime_accounting_t walltime_accounting = nullptr;
815 MembedHolder membedHolder(filenames.size(), filenames.data());
817 /* CAUTION: threads may be started later on in this function, so
818 cr doesn't reflect the final parallel state right now */
821 /* TODO: inputrec should tell us whether we use an algorithm, not a file option */
822 const bool doEssentialDynamics = opt2bSet("-ei", filenames.size(), filenames.data());
823 const bool doRerun = mdrunOptions.rerun;
825 // Handle task-assignment related user options.
826 EmulateGpuNonbonded emulateGpuNonbonded =
827 (getenv("GMX_EMULATE_GPU") != nullptr ? EmulateGpuNonbonded::Yes : EmulateGpuNonbonded::No);
829 std::vector<int> userGpuTaskAssignment;
832 userGpuTaskAssignment = parseUserTaskAssignmentString(hw_opt.userGpuTaskAssignment);
834 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
835 auto nonbondedTarget = findTaskTarget(nbpu_opt);
836 auto pmeTarget = findTaskTarget(pme_opt);
837 auto pmeFftTarget = findTaskTarget(pme_fft_opt);
838 auto bondedTarget = findTaskTarget(bonded_opt);
839 auto updateTarget = findTaskTarget(update_opt);
841 FILE* fplog = nullptr;
842 // If we are appending, we don't write log output because we need
843 // to check that the old log file matches what the checkpoint file
844 // expects. Otherwise, we should start to write log output now if
845 // there is a file ready for it.
846 if (logFileHandle != nullptr && startingBehavior != StartingBehavior::RestartWithAppending)
848 fplog = gmx_fio_getfp(logFileHandle);
850 const bool isSimulationMasterRank = findIsSimulationMasterRank(ms, simulationCommunicator);
851 gmx::LoggerOwner logOwner(buildLogger(fplog, isSimulationMasterRank));
852 gmx::MDLogger mdlog(logOwner.logger());
854 gmx_print_detected_hardware(fplog, isSimulationMasterRank && isMasterSim(ms), mdlog, hwinfo_);
856 std::vector<int> availableDevices =
857 makeListOfAvailableDevices(hwinfo_->deviceInfoList, hw_opt.devicesSelectedByUser);
858 const int numAvailableDevices = gmx::ssize(availableDevices);
860 // Print citation requests after all software/hardware printing
861 pleaseCiteGromacs(fplog);
863 // Note: legacy program logic relies on checking whether these pointers are assigned.
864 // Objects may or may not be allocated later.
865 std::unique_ptr<t_inputrec> inputrec;
866 std::unique_ptr<t_state> globalState;
868 auto partialDeserializedTpr = std::make_unique<PartialDeserializedTprFile>();
870 if (isSimulationMasterRank)
872 // Allocate objects to be initialized by later function calls.
873 /* Only the master rank has the global state */
874 globalState = std::make_unique<t_state>();
875 inputrec = std::make_unique<t_inputrec>();
877 /* Read (nearly) all data required for the simulation
878 * and keep the partly serialized tpr contents to send to other ranks later
880 applyGlobalSimulationState(
881 *inputHolder_.get(), partialDeserializedTpr.get(), globalState.get(), inputrec.get(), &mtop);
884 /* Check and update the hardware options for internal consistency */
885 checkAndUpdateHardwareOptions(
886 mdlog, &hw_opt, isSimulationMasterRank, domdecOptions.numPmeRanks, inputrec.get());
888 if (GMX_THREAD_MPI && isSimulationMasterRank)
890 bool useGpuForNonbonded = false;
891 bool useGpuForPme = false;
894 GMX_RELEASE_ASSERT(inputrec != nullptr, "Keep the compiler happy");
896 // If the user specified the number of ranks, then we must
897 // respect that, but in default mode, we need to allow for
898 // the number of GPUs to choose the number of ranks.
899 auto canUseGpuForNonbonded = buildSupportsNonbondedOnGpu(nullptr);
900 useGpuForNonbonded = decideWhetherToUseGpusForNonbondedWithThreadMpi(
902 numAvailableDevices > 0,
903 userGpuTaskAssignment,
905 canUseGpuForNonbonded,
906 gpuAccelerationOfNonbondedIsUseful(mdlog, *inputrec, GMX_THREAD_MPI),
907 hw_opt.nthreads_tmpi);
908 useGpuForPme = decideWhetherToUseGpusForPmeWithThreadMpi(useGpuForNonbonded,
911 userGpuTaskAssignment,
914 hw_opt.nthreads_tmpi,
915 domdecOptions.numPmeRanks);
917 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
919 /* Determine how many thread-MPI ranks to start.
921 * TODO Over-writing the user-supplied value here does
922 * prevent any possible subsequent checks from working
924 hw_opt.nthreads_tmpi = get_nthreads_mpi(hwinfo_,
932 membedHolder.doMembed());
934 // Now start the threads for thread MPI.
935 spawnThreads(hw_opt.nthreads_tmpi);
936 // The spawned threads enter mdrunner() and execution of
937 // master and spawned threads joins at the end of this block.
940 GMX_RELEASE_ASSERT(!GMX_MPI || ms || simulationCommunicator != MPI_COMM_NULL,
941 "Must have valid communicator unless running a multi-simulation");
942 CommrecHandle crHandle = init_commrec(simulationCommunicator);
943 t_commrec* cr = crHandle.get();
944 GMX_RELEASE_ASSERT(cr != nullptr, "Must have valid commrec");
946 PhysicalNodeCommunicator physicalNodeComm(libraryWorldCommunicator, gmx_physicalnode_id_hash());
948 // If we detected the topology on this system, double-check that it makes sense
949 if (hwinfo_->hardwareTopology->isThisSystem())
951 hardwareTopologyDoubleCheckDetection(mdlog, *hwinfo_->hardwareTopology);
956 /* now broadcast everything to the non-master nodes/threads: */
957 if (!isSimulationMasterRank)
959 // Until now, only the master rank has a non-null pointer.
960 // On non-master ranks, allocate the object that will receive data in the following call.
961 inputrec = std::make_unique<t_inputrec>();
963 init_parallel(cr->mpiDefaultCommunicator,
967 partialDeserializedTpr.get());
969 GMX_RELEASE_ASSERT(inputrec != nullptr, "All ranks should have a valid inputrec now");
970 partialDeserializedTpr.reset(nullptr);
973 !inputrec->useConstantAcceleration,
974 "Linear acceleration has been removed in GROMACS 2022, and was broken for many years "
975 "before that. Use GROMACS 4.5 or earlier if you need this feature.");
977 // Now we decide whether to use the domain decomposition machinery.
978 // Note that this does not necessarily imply actually using multiple domains.
979 // Now the number of ranks is known to all ranks, and each knows
980 // the inputrec read by the master rank. The ranks can now all run
981 // the task-deciding functions and will agree on the result
982 // without needing to communicate.
983 // The LBFGS minimizer, test-particle insertion, normal modes and shell dynamics don't support DD
984 const bool useDomainDecomposition =
985 !(inputrec->eI == IntegrationAlgorithm::LBFGS || EI_TPI(inputrec->eI)
986 || inputrec->eI == IntegrationAlgorithm::NM
987 || gmx_mtop_particletype_count(mtop)[ParticleType::Shell] > 0);
989 // Note that these variables describe only their own node.
991 // Note that when bonded interactions run on a GPU they always run
992 // alongside a nonbonded task, so do not influence task assignment
993 // even though they affect the force calculation workload.
994 bool useGpuForNonbonded = false;
995 bool useGpuForPme = false;
996 bool useGpuForBonded = false;
997 bool useGpuForUpdate = false;
998 bool gpusWereDetected = hwinfo_->ngpu_compatible_tot > 0;
1001 // It's possible that there are different numbers of GPUs on
1002 // different nodes, which is the user's responsibility to
1003 // handle. If unsuitable, we will notice that during task
1005 auto canUseGpuForNonbonded = buildSupportsNonbondedOnGpu(nullptr);
1006 useGpuForNonbonded = decideWhetherToUseGpusForNonbonded(
1008 userGpuTaskAssignment,
1009 emulateGpuNonbonded,
1010 canUseGpuForNonbonded,
1011 gpuAccelerationOfNonbondedIsUseful(mdlog, *inputrec, !GMX_THREAD_MPI),
1013 useGpuForPme = decideWhetherToUseGpusForPme(useGpuForNonbonded,
1015 userGpuTaskAssignment,
1018 cr->sizeOfDefaultCommunicator,
1019 domdecOptions.numPmeRanks,
1021 useGpuForBonded = decideWhetherToUseGpusForBonded(
1022 useGpuForNonbonded, useGpuForPme, bondedTarget, *inputrec, mtop, domdecOptions.numPmeRanks, gpusWereDetected);
1024 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1026 const PmeRunMode pmeRunMode = determinePmeRunMode(useGpuForPme, pmeFftTarget, *inputrec);
1028 // Initialize development feature flags that enabled by environment variable
1029 // and report those features that are enabled.
1030 const DevelopmentFeatureFlags devFlags =
1031 manageDevelopmentFeatures(mdlog, useGpuForNonbonded, pmeRunMode);
1033 const bool useModularSimulator = checkUseModularSimulator(false,
1040 doEssentialDynamics,
1041 membedHolder.doMembed());
1043 ObservablesReducerBuilder observablesReducerBuilder;
1045 // Build restraints.
1046 // TODO: hide restraint implementation details from Mdrunner.
1047 // There is nothing unique about restraints at this point as far as the
1048 // Mdrunner is concerned. The Mdrunner should just be getting a sequence of
1049 // factory functions from the SimulationContext on which to call mdModules_->add().
1050 // TODO: capture all restraints into a single RestraintModule, passed to the runner builder.
1051 for (auto&& restraint : restraintManager_->getRestraints())
1053 auto module = RestraintMDModule::create(restraint, restraint->sites());
1054 mdModules_->add(std::move(module));
1057 // TODO: Error handling
1058 mdModules_->assignOptionsToModules(*inputrec->params, nullptr);
1059 // now that the MDModules know their options, they know which callbacks to sign up to
1060 mdModules_->subscribeToSimulationSetupNotifications();
1061 const auto& setupNotifier = mdModules_->notifiers().simulationSetupNotifier_;
1063 // Notify MdModules of existing logger
1064 setupNotifier.notify(mdlog);
1066 // Notify MdModules of internal parameters, saved into KVT
1067 if (inputrec->internalParameters != nullptr)
1069 setupNotifier.notify(*inputrec->internalParameters);
1072 // Let MdModules know the .tpr filename
1074 gmx::MdRunInputFilename mdRunInputFilename = { ftp2fn(efTPR, filenames.size(), filenames.data()) };
1075 setupNotifier.notify(mdRunInputFilename);
1078 if (fplog != nullptr)
1080 pr_inputrec(fplog, 0, "Input Parameters", inputrec.get(), FALSE);
1081 fprintf(fplog, "\n");
1086 /* In rerun, set velocities to zero if present */
1087 if (doRerun && ((globalState->flags & enumValueToBitMask(StateEntry::V)) != 0))
1089 // rerun does not use velocities
1093 "Rerun trajectory contains velocities. Rerun does only evaluate "
1094 "potential energy and forces. The velocities will be ignored.");
1095 for (int i = 0; i < globalState->natoms; i++)
1097 clear_rvec(globalState->v[i]);
1099 globalState->flags &= ~enumValueToBitMask(StateEntry::V);
1102 /* now make sure the state is initialized and propagated */
1103 set_state_entries(globalState.get(), inputrec.get(), useModularSimulator);
1106 /* NM and TPI parallelize over force/energy calculations, not atoms,
1107 * so we need to initialize and broadcast the global state.
1109 if (inputrec->eI == IntegrationAlgorithm::NM || inputrec->eI == IntegrationAlgorithm::TPI)
1113 globalState = std::make_unique<t_state>();
1115 broadcastStateWithoutDynamics(
1116 cr->mpiDefaultCommunicator, haveDDAtomOrdering(*cr), PAR(cr), globalState.get());
1119 /* A parallel command line option consistency check that we can
1120 only do after any threads have started. */
1122 && (domdecOptions.numCells[XX] > 1 || domdecOptions.numCells[YY] > 1
1123 || domdecOptions.numCells[ZZ] > 1 || domdecOptions.numPmeRanks > 0))
1126 "The -dd or -npme option request a parallel simulation, "
1128 "but %s was compiled without threads or MPI enabled",
1129 output_env_get_program_display_name(oenv));
1130 #elif GMX_THREAD_MPI
1131 "but the number of MPI-threads (option -ntmpi) is not set or is 1");
1133 "but %s was not started through mpirun/mpiexec or only one rank was requested "
1134 "through mpirun/mpiexec",
1135 output_env_get_program_display_name(oenv));
1139 if (doRerun && (EI_ENERGY_MINIMIZATION(inputrec->eI) || IntegrationAlgorithm::NM == inputrec->eI))
1142 "The .mdp file specified an energy mininization or normal mode algorithm, and "
1143 "these are not compatible with mdrun -rerun");
1146 /* NMR restraints must be initialized before load_checkpoint,
1147 * since with time averaging the history is added to t_state.
1148 * For proper consistency check we therefore need to extend
1150 * So the PME-only nodes (if present) will also initialize
1151 * the distance restraints.
1154 /* This needs to be called before read_checkpoint to extend the state */
1155 t_disresdata* disresdata;
1156 snew(disresdata, 1);
1160 DisResRunMode::MDRun,
1161 MASTER(cr) ? DDRole::Master : DDRole::Agent,
1162 PAR(cr) ? NumRanks::Multiple : NumRanks::Single,
1167 replExParams.exchangeInterval > 0);
1169 if (gmx_mtop_ftype_count(mtop, F_ORIRES) > 0 && isSimulationMasterRank)
1171 extendStateWithOriresHistory(mtop, *inputrec, globalState.get());
1174 auto deform = prepareBoxDeformation(globalState != nullptr ? globalState->box : box,
1175 MASTER(cr) ? DDRole::Master : DDRole::Agent,
1176 PAR(cr) ? NumRanks::Multiple : NumRanks::Single,
1177 cr->mpi_comm_mygroup,
1181 /* We have to remember the generation's first step before reading checkpoint.
1182 This way, we can report to the F@H core both the generation's first step
1183 and the restored first step, thus making it able to distinguish between
1184 an interruption/resume and start of the n-th generation simulation.
1185 Having this information, the F@H core can correctly calculate and report
1188 int gen_first_step = 0;
1191 gen_first_step = inputrec->init_step;
1195 ObservablesHistory observablesHistory = {};
1197 auto modularSimulatorCheckpointData = std::make_unique<ReadCheckpointDataHolder>();
1198 if (startingBehavior != StartingBehavior::NewSimulation)
1200 /* Check if checkpoint file exists before doing continuation.
1201 * This way we can use identical input options for the first and subsequent runs...
1203 if (mdrunOptions.numStepsCommandline > -2)
1205 /* Temporarily set the number of steps to unlimited to avoid
1206 * triggering the nsteps check in load_checkpoint().
1207 * This hack will go away soon when the -nsteps option is removed.
1209 inputrec->nsteps = -1;
1212 // Finish applying initial simulation state information from external sources on all ranks.
1213 // Reconcile checkpoint file data with Mdrunner state established up to this point.
1214 applyLocalState(*inputHolder_.get(),
1217 domdecOptions.numCells,
1220 &observablesHistory,
1221 mdrunOptions.reproducible,
1222 mdModules_->notifiers(),
1223 modularSimulatorCheckpointData.get(),
1224 useModularSimulator);
1225 // TODO: (#3652) Synchronize filesystem state, SimulationInput contents, and program
1227 // on all code paths.
1228 // Write checkpoint or provide hook to update SimulationInput.
1229 // If there was a checkpoint file, SimulationInput contains more information
1230 // than if there wasn't. At this point, we have synchronized the in-memory
1231 // state with the filesystem state only for restarted simulations. We should
1232 // be calling applyLocalState unconditionally and expect that the completeness
1233 // of SimulationInput is not dependent on its creation method.
1235 if (startingBehavior == StartingBehavior::RestartWithAppending && logFileHandle)
1237 // Now we can start normal logging to the truncated log file.
1238 fplog = gmx_fio_getfp(logFileHandle);
1239 prepareLogAppending(fplog);
1240 logOwner = buildLogger(fplog, MASTER(cr));
1241 mdlog = logOwner.logger();
1248 fcRegisterSteps(inputrec->nsteps + inputrec->init_step, gen_first_step);
1252 if (mdrunOptions.numStepsCommandline > -2)
1257 "The -nsteps functionality is deprecated, and may be removed in a future "
1259 "Consider using gmx convert-tpr -nsteps or changing the appropriate .mdp "
1262 /* override nsteps with value set on the commandline */
1263 override_nsteps_cmdline(mdlog, mdrunOptions.numStepsCommandline, inputrec.get());
1265 if (isSimulationMasterRank)
1267 copy_mat(globalState->box, box);
1272 gmx_bcast(sizeof(box), box, cr->mpiDefaultCommunicator);
1275 if (inputrec->cutoff_scheme != CutoffScheme::Verlet)
1278 "This group-scheme .tpr file can no longer be run by mdrun. Please update to the "
1279 "Verlet scheme, or use an earlier version of GROMACS if necessary.");
1281 /* Update rlist and nstlist. */
1282 /* Note: prepare_verlet_scheme is calling increaseNstlist(...), which (while attempting to
1283 * increase rlist) tries to check if the newly chosen value fits with the DD scheme. As this is
1284 * run before any DD scheme is set up, this check is never executed. See #3334 for more details.
1286 prepare_verlet_scheme(fplog,
1292 useGpuForNonbonded || (emulateGpuNonbonded == EmulateGpuNonbonded::Yes),
1295 // We need to decide on update groups early, as this affects
1296 // inter-domain communication distances.
1297 auto updateGroupingsPerMoleculeType = makeUpdateGroupingsPerMoleculeType(mtop);
1298 const real maxUpdateGroupRadius = computeMaxUpdateGroupRadius(
1299 mtop, updateGroupingsPerMoleculeType, maxReferenceTemperature(*inputrec));
1300 const real cutoffMargin = std::sqrt(max_cutoff2(inputrec->pbcType, box)) - inputrec->rlist;
1301 UpdateGroups updateGroups = makeUpdateGroups(mdlog,
1302 std::move(updateGroupingsPerMoleculeType),
1303 maxUpdateGroupRadius,
1304 useDomainDecomposition,
1305 systemHasConstraintsOrVsites(mtop),
1310 const bool haveFrozenAtoms = inputrecFrozenAtoms(inputrec.get());
1312 useGpuForUpdate = decideWhetherToUseGpuForUpdate(useDomainDecomposition,
1313 updateGroups.useUpdateGroups(),
1315 domdecOptions.numPmeRanks > 0,
1321 doEssentialDynamics,
1322 gmx_mtop_ftype_count(mtop, F_ORIRES) > 0,
1328 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1330 bool useGpuDirectHalo = false;
1332 if (useGpuForNonbonded)
1334 // cr->npmenodes is not yet initialized.
1335 // domdecOptions.numPmeRanks == -1 results in 0 separate PME ranks when useGpuForNonbonded is true.
1336 // Todo: remove this assumption later once auto mode has support for separate PME rank
1337 const int numPmeRanks = domdecOptions.numPmeRanks > 0 ? domdecOptions.numPmeRanks : 0;
1338 bool havePPDomainDecomposition = (cr->sizeOfDefaultCommunicator - numPmeRanks) > 1;
1339 useGpuDirectHalo = decideWhetherToUseGpuForHalo(devFlags,
1340 havePPDomainDecomposition,
1342 useModularSimulator,
1344 EI_ENERGY_MINIMIZATION(inputrec->eI));
1347 // This builder is necessary while we have multi-part construction
1348 // of DD. Before DD is constructed, we use the existence of
1349 // the builder object to indicate that further construction of DD
1351 std::unique_ptr<DomainDecompositionBuilder> ddBuilder;
1352 if (useDomainDecomposition)
1354 // P2P GPU comm + GPU update leads to case in which we enqueue async work for multiple
1355 // timesteps. DLB needs to be disabled in that case
1356 const bool directGpuCommUsedWithGpuUpdate = GMX_THREAD_MPI && useGpuDirectHalo && useGpuForUpdate;
1357 ddBuilder = std::make_unique<DomainDecompositionBuilder>(
1364 mdModules_->notifiers(),
1366 updateGroups.updateGroupingPerMoleculeType(),
1367 updateGroups.useUpdateGroups(),
1368 updateGroups.maxUpdateGroupRadius(),
1369 positionsFromStatePointer(globalState.get()),
1372 directGpuCommUsedWithGpuUpdate);
1376 /* PME, if used, is done on all nodes with 1D decomposition */
1377 cr->nnodes = cr->sizeOfDefaultCommunicator;
1378 cr->sim_nodeid = cr->rankInDefaultCommunicator;
1379 cr->nodeid = cr->rankInDefaultCommunicator;
1381 cr->duty = (DUTY_PP | DUTY_PME);
1383 if (inputrec->pbcType == PbcType::Screw)
1385 gmx_fatal(FARGS, "pbc=screw is only implemented with domain decomposition");
1389 // Produce the task assignment for this rank - done after DD is constructed
1390 GpuTaskAssignments gpuTaskAssignments = GpuTaskAssignmentsBuilder::build(
1392 userGpuTaskAssignment,
1394 simulationCommunicator,
1402 thisRankHasDuty(cr, DUTY_PP),
1403 // TODO cr->duty & DUTY_PME should imply that a PME
1404 // algorithm is active, but currently does not.
1405 EEL_PME(inputrec->coulombtype) && thisRankHasDuty(cr, DUTY_PME));
1407 // Get the device handles for the modules, nullptr when no task is assigned.
1409 DeviceInformation* deviceInfo = gpuTaskAssignments.initDevice(&deviceId);
1411 // timing enabling - TODO put this in gpu_utils (even though generally this is just option handling?)
1412 bool useTiming = true;
1416 /* WARNING: CUDA timings are incorrect with multiple streams.
1417 * This is the main reason why they are disabled by default.
1419 // TODO: Consider turning on by default when we can detect nr of streams.
1420 useTiming = (getenv("GMX_ENABLE_GPU_TIMING") != nullptr);
1422 else if (GMX_GPU_OPENCL)
1424 useTiming = (getenv("GMX_DISABLE_GPU_TIMING") == nullptr);
1427 // TODO Currently this is always built, yet DD partition code
1428 // checks if it is built before using it. Probably it should
1429 // become an MDModule that is made only when another module
1430 // requires it (e.g. pull, CompEl, density fitting), so that we
1431 // don't update the local atom sets unilaterally every step.
1432 LocalAtomSetManager atomSets;
1434 // Local state and topology are declared (and perhaps constructed)
1435 // now, because DD needs them for the LocalTopologyChecker, but
1436 // they do not contain valid data until after the first DD
1438 std::unique_ptr<t_state> localStateInstance;
1439 t_state* localState;
1440 gmx_localtop_t localTopology(mtop.ffparams);
1444 localStateInstance = std::make_unique<t_state>();
1445 localState = localStateInstance.get();
1446 // TODO Pass the GPU streams to ddBuilder to use in buffer
1447 // transfers (e.g. halo exchange)
1448 cr->dd = ddBuilder->build(&atomSets, localTopology, *localState, &observablesReducerBuilder);
1449 // The builder's job is done, so destruct it
1450 ddBuilder.reset(nullptr);
1451 // Note that local state still does not exist yet.
1455 // Without DD, the local state is merely an alias to the global state,
1456 // so we don't need to allocate anything.
1457 localState = globalState.get();
1460 // Ensure that all atoms within the same update group are in the
1461 // same periodic image. Otherwise, a simulation that did not use
1462 // update groups (e.g. a single-rank simulation) cannot always be
1463 // correctly restarted in a way that does use update groups
1464 // (e.g. a multi-rank simulation).
1465 if (isSimulationMasterRank)
1467 const bool useUpdateGroups = cr->dd ? ddUsesUpdateGroups(*cr->dd) : false;
1468 if (useUpdateGroups)
1470 putUpdateGroupAtomsInSamePeriodicImage(*cr->dd, mtop, globalState->box, globalState->x);
1474 const bool printHostName = (cr->nnodes > 1);
1475 gpuTaskAssignments.reportGpuUsage(mdlog, printHostName, useGpuForBonded, pmeRunMode, useGpuForUpdate);
1477 const bool disableNonbondedCalculation = (getenv("GMX_NO_NONBONDED") != nullptr);
1478 if (disableNonbondedCalculation)
1480 /* turn off non-bonded calculations */
1481 GMX_LOG(mdlog.warning)
1484 "Found environment variable GMX_NO_NONBONDED.\n"
1485 "Disabling nonbonded calculations.");
1488 MdrunScheduleWorkload runScheduleWork;
1490 // Also populates the simulation constant workload description.
1491 // Note: currently the default duty is DUTY_PP | DUTY_PME for all simulations, including those without PME,
1492 // so this boolean is sufficient on all ranks to determine whether separate PME ranks are used,
1493 // but this will no longer be the case if cr->duty is changed for !EEL_PME(fr->ic->eeltype).
1494 const bool haveSeparatePmeRank = (!thisRankHasDuty(cr, DUTY_PP) || !thisRankHasDuty(cr, DUTY_PME));
1495 runScheduleWork.simulationWork = createSimulationWorkload(*inputrec,
1496 disableNonbondedCalculation,
1498 havePPDomainDecomposition(cr),
1499 haveSeparatePmeRank,
1506 std::unique_ptr<DeviceStreamManager> deviceStreamManager = nullptr;
1508 if (deviceInfo != nullptr)
1510 if (runScheduleWork.simulationWork.havePpDomainDecomposition && thisRankHasDuty(cr, DUTY_PP))
1512 dd_setup_dlb_resource_sharing(cr, deviceId);
1514 deviceStreamManager = std::make_unique<DeviceStreamManager>(
1515 *deviceInfo, havePPDomainDecomposition(cr), runScheduleWork.simulationWork, useTiming);
1518 // If the user chose a task assignment, give them some hints
1519 // where appropriate.
1520 if (!userGpuTaskAssignment.empty())
1522 gpuTaskAssignments.logPerformanceHints(mdlog, numAvailableDevices);
1527 /* After possible communicator splitting in make_dd_communicators.
1528 * we can set up the intra/inter node communication.
1530 gmx_setup_nodecomm(fplog, cr);
1536 GMX_LOG(mdlog.warning)
1538 .appendTextFormatted(
1539 "This is simulation %d out of %d running as a composite GROMACS\n"
1540 "multi-simulation job. Setup for this simulation:\n",
1541 ms->simulationIndex_,
1542 ms->numSimulations_);
1544 GMX_LOG(mdlog.warning)
1545 .appendTextFormatted("Using %d MPI %s\n",
1548 cr->nnodes == 1 ? "thread" : "threads"
1550 cr->nnodes == 1 ? "process" : "processes"
1556 // If mdrun -pin auto honors any affinity setting that already
1557 // exists. If so, it is nice to provide feedback about whether
1558 // that existing affinity setting was from OpenMP or something
1559 // else, so we run this code both before and after we initialize
1560 // the OpenMP support.
1561 gmx_check_thread_affinity_set(mdlog, &hw_opt, hwinfo_->nthreads_hw_avail, FALSE);
1562 /* Check and update the number of OpenMP threads requested */
1563 checkAndUpdateRequestedNumOpenmpThreads(
1564 &hw_opt, *hwinfo_, cr, ms, physicalNodeComm.size_, pmeRunMode, mtop, *inputrec);
1566 gmx_omp_nthreads_init(mdlog,
1568 hwinfo_->nthreads_hw_avail,
1569 physicalNodeComm.size_,
1570 hw_opt.nthreads_omp,
1571 hw_opt.nthreads_omp_pme,
1572 !thisRankHasDuty(cr, DUTY_PP));
1574 const bool bEnableFPE = gmxShouldEnableFPExceptions();
1575 // FIXME - reconcile with gmx_feenableexcept() call from CommandLineModuleManager::run()
1578 gmx_feenableexcept();
1581 /* Now that we know the setup is consistent, check for efficiency */
1582 check_resource_division_efficiency(
1583 hwinfo_, gpuTaskAssignments.thisRankHasAnyGpuTask(), mdrunOptions.ntompOptionIsSet, cr, mdlog);
1585 /* getting number of PP/PME threads on this MPI / tMPI rank.
1586 PME: env variable should be read only on one node to make sure it is
1587 identical everywhere;
1589 const int numThreadsOnThisRank = thisRankHasDuty(cr, DUTY_PP)
1590 ? gmx_omp_nthreads_get(ModuleMultiThread::Nonbonded)
1591 : gmx_omp_nthreads_get(ModuleMultiThread::Pme);
1592 checkHardwareOversubscription(
1593 numThreadsOnThisRank, cr->nodeid, *hwinfo_->hardwareTopology, physicalNodeComm, mdlog);
1595 // Enable Peer access between GPUs where available
1596 // Only for DD, only master PP rank needs to perform setup, and only if thread MPI plus
1597 // any of the GPU communication features are active.
1598 if (haveDDAtomOrdering(*cr) && MASTER(cr) && thisRankHasDuty(cr, DUTY_PP) && GMX_THREAD_MPI
1599 && (runScheduleWork.simulationWork.useGpuHaloExchange
1600 || runScheduleWork.simulationWork.useGpuPmePpCommunication))
1602 setupGpuDevicePeerAccess(gpuTaskAssignments.deviceIdsAssigned(), mdlog);
1605 if (hw_opt.threadAffinity != ThreadAffinity::Off)
1607 /* Before setting affinity, check whether the affinity has changed
1608 * - which indicates that probably the OpenMP library has changed it
1609 * since we first checked).
1611 gmx_check_thread_affinity_set(mdlog, &hw_opt, hwinfo_->nthreads_hw_avail, TRUE);
1613 int numThreadsOnThisNode, intraNodeThreadOffset;
1614 analyzeThreadsOnThisNode(
1615 physicalNodeComm, numThreadsOnThisRank, &numThreadsOnThisNode, &intraNodeThreadOffset);
1617 /* Set the CPU affinity */
1618 gmx_set_thread_affinity(mdlog,
1621 *hwinfo_->hardwareTopology,
1622 numThreadsOnThisRank,
1623 numThreadsOnThisNode,
1624 intraNodeThreadOffset,
1628 if (mdrunOptions.timingOptions.resetStep > -1)
1633 "The -resetstep functionality is deprecated, and may be removed in a "
1636 std::unique_ptr<gmx_wallcycle> wcycle =
1637 wallcycle_init(fplog, mdrunOptions.timingOptions.resetStep, cr);
1641 /* Master synchronizes its value of reset_counters with all nodes
1642 * including PME only nodes */
1643 int64_t reset_counters = wcycle_get_reset_counters(wcycle.get());
1644 gmx_bcast(sizeof(reset_counters), &reset_counters, cr->mpi_comm_mysim);
1645 wcycle_set_reset_counters(wcycle.get(), reset_counters);
1648 // Membrane embedding must be initialized before we call init_forcerec()
1649 membedHolder.initializeMembed(fplog,
1656 &mdrunOptions.checkpointOptions.period);
1658 const bool thisRankHasPmeGpuTask = gpuTaskAssignments.thisRankHasPmeGpuTask();
1659 std::unique_ptr<MDAtoms> mdAtoms;
1660 std::unique_ptr<VirtualSitesHandler> vsite;
1663 if (thisRankHasDuty(cr, DUTY_PP))
1665 setupNotifier.notify(*cr);
1666 setupNotifier.notify(&atomSets);
1667 setupNotifier.notify(mtop);
1668 setupNotifier.notify(inputrec->pbcType);
1669 setupNotifier.notify(SimulationTimeStep{ inputrec->delta_t });
1670 /* Initiate forcerecord */
1671 fr = std::make_unique<t_forcerec>();
1672 fr->forceProviders = mdModules_->initForceProviders();
1673 init_forcerec(fplog,
1675 runScheduleWork.simulationWork,
1681 opt2fn("-table", filenames.size(), filenames.data()),
1682 opt2fn("-tablep", filenames.size(), filenames.data()),
1683 opt2fns("-tableb", filenames.size(), filenames.data()),
1685 // Dirty hack, for fixing disres and orires should be made mdmodules
1686 fr->fcdata->disres = disresdata;
1687 if (gmx_mtop_ftype_count(mtop, F_ORIRES) > 0)
1689 fr->fcdata->orires = std::make_unique<t_oriresdata>(
1690 fplog, mtop, *inputrec, ms, globalState.get(), &atomSets);
1693 // Save a handle to device stream manager to use elsewhere in the code
1694 // TODO: Forcerec is not a correct place to store it.
1695 fr->deviceStreamManager = deviceStreamManager.get();
1697 if (runScheduleWork.simulationWork.useGpuPmePpCommunication && !thisRankHasDuty(cr, DUTY_PME))
1700 deviceStreamManager != nullptr,
1701 "GPU device stream manager should be valid in order to use PME-PP direct "
1704 deviceStreamManager->streamIsValid(DeviceStreamType::PmePpTransfer),
1705 "GPU PP-PME stream should be valid in order to use GPU PME-PP direct "
1707 fr->pmePpCommGpu = std::make_unique<gmx::PmePpCommGpu>(
1710 &cr->dd->pmeForceReceiveBuffer,
1711 deviceStreamManager->context(),
1712 deviceStreamManager->stream(DeviceStreamType::PmePpTransfer));
1715 fr->nbv = Nbnxm::init_nb_verlet(mdlog,
1720 runScheduleWork.simulationWork.useGpuNonbonded,
1721 deviceStreamManager.get(),
1725 // TODO: Move the logic below to a GPU bonded builder
1726 if (runScheduleWork.simulationWork.useGpuBonded)
1728 GMX_RELEASE_ASSERT(deviceStreamManager != nullptr,
1729 "GPU device stream manager should be valid in order to use GPU "
1730 "version of bonded forces.");
1731 fr->listedForcesGpu = std::make_unique<ListedForcesGpu>(
1733 fr->ic->epsfac * fr->fudgeQQ,
1734 deviceStreamManager->context(),
1735 deviceStreamManager->bondedStream(havePPDomainDecomposition(cr)),
1738 fr->longRangeNonbondeds = std::make_unique<CpuPpLongRangeNonbondeds>(fr->n_tpi,
1739 fr->ic->ewaldcoeff_q,
1749 /* Initialize the mdAtoms structure.
1750 * mdAtoms is not filled with atom data,
1751 * as this can not be done now with domain decomposition.
1753 mdAtoms = makeMDAtoms(fplog, mtop, *inputrec, thisRankHasPmeGpuTask);
1754 if (globalState && thisRankHasPmeGpuTask)
1756 // The pinning of coordinates in the global state object works, because we only use
1757 // PME on GPU without DD or on a separate PME rank, and because the local state pointer
1758 // points to the global state object without DD.
1759 // FIXME: MD and EM separately set up the local state - this should happen in the same
1760 // function, which should also perform the pinning.
1761 changePinningPolicy(&globalState->x, pme_get_pinning_policy());
1764 /* Initialize the virtual site communication */
1765 vsite = makeVirtualSitesHandler(
1766 mtop, cr, fr->pbcType, updateGroups.updateGroupingPerMoleculeType());
1768 calc_shifts(box, fr->shift_vec);
1770 /* With periodic molecules the charge groups should be whole at start up
1771 * and the virtual sites should not be far from their proper positions.
1773 if (!inputrec->bContinuation && MASTER(cr)
1774 && !(inputrec->pbcType != PbcType::No && inputrec->bPeriodicMols))
1776 /* Make molecules whole at start of run */
1777 if (fr->pbcType != PbcType::No)
1779 do_pbc_first_mtop(fplog, inputrec->pbcType, box, &mtop, globalState->x.rvec_array());
1783 /* Correct initial vsite positions are required
1784 * for the initial distribution in the domain decomposition
1785 * and for the initial shell prediction.
1787 constructVirtualSitesGlobal(mtop, globalState->x);
1790 // Make the DD reverse topology, now that any vsites that are present are available
1791 if (haveDDAtomOrdering(*cr))
1793 dd_make_reverse_top(fplog, cr->dd, mtop, vsite.get(), *inputrec, domdecOptions.ddBondedChecking);
1796 if (EEL_PME(fr->ic->eeltype) || EVDW_PME(fr->ic->vdwtype))
1798 ewaldcoeff_q = fr->ic->ewaldcoeff_q;
1799 ewaldcoeff_lj = fr->ic->ewaldcoeff_lj;
1804 /* This is a PME only node */
1806 GMX_ASSERT(globalState == nullptr,
1807 "We don't need the state on a PME only rank and expect it to be unitialized");
1809 ewaldcoeff_q = calc_ewaldcoeff_q(inputrec->rcoulomb, inputrec->ewald_rtol);
1810 ewaldcoeff_lj = calc_ewaldcoeff_lj(inputrec->rvdw, inputrec->ewald_rtol_lj);
1813 gmx_pme_t* sepPmeData = nullptr;
1814 // This reference hides the fact that PME data is owned by runner on PME-only ranks and by forcerec on other ranks
1815 GMX_ASSERT(thisRankHasDuty(cr, DUTY_PP) == (fr != nullptr),
1816 "Double-checking that only PME-only ranks have no forcerec");
1817 gmx_pme_t*& pmedata = fr ? fr->pmedata : sepPmeData;
1819 // TODO should live in ewald module once its testing is improved
1821 // Later, this program could contain kernels that might be later
1822 // re-used as auto-tuning progresses, or subsequent simulations
1824 PmeGpuProgramStorage pmeGpuProgram;
1825 if (thisRankHasPmeGpuTask)
1828 (deviceStreamManager != nullptr),
1829 "GPU device stream manager should be initialized in order to use GPU for PME.");
1830 GMX_RELEASE_ASSERT((deviceInfo != nullptr),
1831 "GPU device should be initialized in order to use GPU for PME.");
1832 pmeGpuProgram = buildPmeGpuProgram(deviceStreamManager->context());
1835 /* Initiate PME if necessary,
1836 * either on all nodes or on dedicated PME nodes only. */
1837 if (EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype))
1839 if (mdAtoms && mdAtoms->mdatoms())
1841 nChargePerturbed = mdAtoms->mdatoms()->nChargePerturbed;
1842 if (EVDW_PME(inputrec->vdwtype))
1844 nTypePerturbed = mdAtoms->mdatoms()->nTypePerturbed;
1847 if (cr->npmenodes > 0)
1849 /* The PME only nodes need to know nChargePerturbed(FEP on Q) and nTypePerturbed(FEP on LJ)*/
1850 gmx_bcast(sizeof(nChargePerturbed), &nChargePerturbed, cr->mpi_comm_mysim);
1851 gmx_bcast(sizeof(nTypePerturbed), &nTypePerturbed, cr->mpi_comm_mysim);
1854 if (thisRankHasDuty(cr, DUTY_PME))
1858 // TODO: This should be in the builder.
1859 GMX_RELEASE_ASSERT(!runScheduleWork.simulationWork.useGpuPme
1860 || (deviceStreamManager != nullptr),
1861 "Device stream manager should be valid in order to use GPU "
1864 !runScheduleWork.simulationWork.useGpuPme
1865 || deviceStreamManager->streamIsValid(DeviceStreamType::Pme),
1866 "GPU PME stream should be valid in order to use GPU version of PME.");
1868 const DeviceContext* deviceContext = runScheduleWork.simulationWork.useGpuPme
1869 ? &deviceStreamManager->context()
1871 const DeviceStream* pmeStream =
1872 runScheduleWork.simulationWork.useGpuPme
1873 ? &deviceStreamManager->stream(DeviceStreamType::Pme)
1876 pmedata = gmx_pme_init(cr,
1877 getNumPmeDomains(cr->dd),
1879 nChargePerturbed != 0,
1880 nTypePerturbed != 0,
1881 mdrunOptions.reproducible,
1884 gmx_omp_nthreads_get(ModuleMultiThread::Pme),
1889 pmeGpuProgram.get(),
1892 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1897 if (EI_DYNAMICS(inputrec->eI))
1899 /* Turn on signal handling on all nodes */
1901 * (A user signal from the PME nodes (if any)
1902 * is communicated to the PP nodes.
1904 signal_handler_install();
1907 pull_t* pull_work = nullptr;
1908 if (thisRankHasDuty(cr, DUTY_PP))
1910 /* Assumes uniform use of the number of OpenMP threads */
1911 walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(ModuleMultiThread::Default));
1913 if (inputrec->bPull)
1915 /* Initialize pull code */
1916 pull_work = init_pull(fplog,
1917 inputrec->pull.get(),
1922 inputrec->fepvals->init_lambda);
1923 if (inputrec->pull->bXOutAverage || inputrec->pull->bFOutAverage)
1925 initPullHistory(pull_work, &observablesHistory);
1927 if (EI_DYNAMICS(inputrec->eI) && MASTER(cr))
1929 init_pull_output_files(pull_work, filenames.size(), filenames.data(), oenv, startingBehavior);
1933 std::unique_ptr<EnforcedRotation> enforcedRotation;
1936 /* Initialize enforced rotation code */
1937 enforcedRotation = init_rot(fplog,
1950 t_swap* swap = nullptr;
1951 if (inputrec->eSwapCoords != SwapType::No)
1953 /* Initialize ion swapping code */
1954 swap = init_swapcoords(fplog,
1956 opt2fn_master("-swap", filenames.size(), filenames.data(), cr),
1959 &observablesHistory,
1967 /* Let makeConstraints know whether we have essential dynamics constraints. */
1968 auto constr = makeConstraints(mtop,
1971 doEssentialDynamics,
1974 updateGroups.useUpdateGroups(),
1979 &observablesReducerBuilder);
1981 /* Energy terms and groups */
1982 gmx_enerdata_t enerd(mtop.groups.groups[SimulationAtomGroupType::EnergyOutput].size(),
1983 inputrec->fepvals->n_lambda);
1985 // cos acceleration is only supported by md, but older tpr
1986 // files might still combine it with other integrators
1987 GMX_RELEASE_ASSERT(inputrec->cos_accel == 0.0 || inputrec->eI == IntegrationAlgorithm::MD,
1988 "cos_acceleration is only supported by integrator=md");
1990 /* Kinetic energy data */
1991 gmx_ekindata_t ekind(inputrec->opts.ngtc,
1992 inputrec->cos_accel,
1993 gmx_omp_nthreads_get(ModuleMultiThread::Update));
1995 /* Set up interactive MD (IMD) */
1996 auto imdSession = makeImdSession(inputrec.get(),
2003 MASTER(cr) ? globalState->x : gmx::ArrayRef<gmx::RVec>(),
2007 mdrunOptions.imdOptions,
2010 if (haveDDAtomOrdering(*cr))
2012 GMX_RELEASE_ASSERT(fr, "fr was NULL while cr->duty was DUTY_PP");
2013 /* This call is not included in init_domain_decomposition
2014 * because fr->atomInfoForEachMoleculeBlock is set later.
2016 makeBondedLinks(cr->dd, mtop, fr->atomInfoForEachMoleculeBlock);
2019 if (runScheduleWork.simulationWork.useGpuBufferOps)
2021 fr->gpuForceReduction[gmx::AtomLocality::Local] = std::make_unique<gmx::GpuForceReduction>(
2022 deviceStreamManager->context(),
2023 deviceStreamManager->stream(gmx::DeviceStreamType::NonBondedLocal),
2025 fr->gpuForceReduction[gmx::AtomLocality::NonLocal] = std::make_unique<gmx::GpuForceReduction>(
2026 deviceStreamManager->context(),
2027 deviceStreamManager->stream(gmx::DeviceStreamType::NonBondedNonLocal),
2031 std::unique_ptr<gmx::StatePropagatorDataGpu> stateGpu;
2032 if (gpusWereDetected
2033 && ((runScheduleWork.simulationWork.useGpuPme && thisRankHasDuty(cr, DUTY_PME))
2034 || runScheduleWork.simulationWork.useGpuBufferOps))
2036 GpuApiCallBehavior transferKind =
2037 (inputrec->eI == IntegrationAlgorithm::MD && !doRerun && !useModularSimulator)
2038 ? GpuApiCallBehavior::Async
2039 : GpuApiCallBehavior::Sync;
2040 GMX_RELEASE_ASSERT(deviceStreamManager != nullptr,
2041 "GPU device stream manager should be initialized to use GPU.");
2042 stateGpu = std::make_unique<gmx::StatePropagatorDataGpu>(
2043 *deviceStreamManager, transferKind, pme_gpu_get_block_size(fr->pmedata), wcycle.get());
2044 fr->stateGpu = stateGpu.get();
2047 GMX_ASSERT(stopHandlerBuilder_, "Runner must provide StopHandlerBuilder to simulator.");
2048 SimulatorBuilder simulatorBuilder;
2050 simulatorBuilder.add(SimulatorStateData(
2051 globalState.get(), localState, &observablesHistory, &enerd, &ekind));
2052 simulatorBuilder.add(std::move(membedHolder));
2053 simulatorBuilder.add(std::move(stopHandlerBuilder_));
2054 simulatorBuilder.add(SimulatorConfig(mdrunOptions, startingBehavior, &runScheduleWork));
2057 simulatorBuilder.add(SimulatorEnv(fplog, cr, ms, mdlog, oenv, &observablesReducerBuilder));
2058 simulatorBuilder.add(Profiling(&nrnb, walltime_accounting, wcycle.get()));
2059 simulatorBuilder.add(ConstraintsParam(
2060 constr.get(), enforcedRotation ? enforcedRotation->getLegacyEnfrot() : nullptr, vsite.get()));
2061 // TODO: Separate `fr` to a separate add, and make the `build` handle the coupling sensibly.
2062 simulatorBuilder.add(LegacyInput(
2063 static_cast<int>(filenames.size()), filenames.data(), inputrec.get(), fr.get()));
2064 simulatorBuilder.add(ReplicaExchangeParameters(replExParams));
2065 simulatorBuilder.add(InteractiveMD(imdSession.get()));
2066 simulatorBuilder.add(SimulatorModules(mdModules_->outputProvider(), mdModules_->notifiers()));
2067 simulatorBuilder.add(CenterOfMassPulling(pull_work));
2068 // Todo move to an MDModule
2069 simulatorBuilder.add(IonSwapping(swap));
2070 simulatorBuilder.add(TopologyData(mtop, &localTopology, mdAtoms.get()));
2071 simulatorBuilder.add(BoxDeformationHandle(deform.get()));
2072 simulatorBuilder.add(std::move(modularSimulatorCheckpointData));
2074 // build and run simulator object based on user-input
2075 auto simulator = simulatorBuilder.build(useModularSimulator);
2078 if (fr->pmePpCommGpu)
2080 // destroy object since it is no longer required. (This needs to be done while the GPU context still exists.)
2081 fr->pmePpCommGpu.reset();
2084 if (inputrec->bPull)
2086 finish_pull(pull_work);
2088 finish_swapcoords(swap);
2092 GMX_RELEASE_ASSERT(pmedata, "pmedata was NULL while cr->duty was not DUTY_PP");
2094 walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(ModuleMultiThread::Pme));
2095 gmx_pmeonly(pmedata,
2099 walltime_accounting,
2102 runScheduleWork.simulationWork.useGpuPmePpCommunication,
2103 deviceStreamManager.get());
2106 wallcycle_stop(wcycle.get(), WallCycleCounter::Run);
2108 /* Finish up, write some stuff
2109 * if rerunMD, don't write last frame again
2117 walltime_accounting,
2118 fr ? fr->nbv.get() : nullptr,
2120 EI_DYNAMICS(inputrec->eI) && !isMultiSim(ms));
2123 deviceStreamManager.reset(nullptr);
2127 gmx_pme_destroy(pmedata);
2131 // FIXME: this is only here to manually unpin mdAtoms->chargeA_ and state->x,
2132 // before we destroy the GPU context(s)
2133 // Pinned buffers are associated with contexts in CUDA.
2134 // As soon as we destroy GPU contexts after mdrunner() exits, these lines should go.
2135 mdAtoms.reset(nullptr);
2136 globalState.reset(nullptr);
2137 localStateInstance.reset(nullptr);
2138 mdModules_.reset(nullptr); // destruct force providers here as they might also use the GPU
2139 fr.reset(nullptr); // destruct forcerec before gpu
2140 // TODO convert to C++ so we can get rid of these frees
2143 if (!hwinfo_->deviceInfoList.empty())
2145 /* stop the GPU profiler (only CUDA) */
2149 /* With tMPI we need to wait for all ranks to finish deallocation before
2150 * destroying the CUDA context as some tMPI ranks may be sharing
2153 * This is not a concern in OpenCL where we use one context per rank.
2155 * Note: it is safe to not call the barrier on the ranks which do not use GPU,
2156 * but it is easier and more futureproof to call it on the whole node.
2158 * Note that this function needs to be called even if GPUs are not used
2159 * in this run because the PME ranks have no knowledge of whether GPUs
2160 * are used or not, but all ranks need to enter the barrier below.
2161 * \todo Remove this physical node barrier after making sure
2162 * that it's not needed anymore (with a shared GPU run).
2166 physicalNodeComm.barrier();
2169 if (!devFlags.usingCudaAwareMpi)
2171 // Don't reset GPU in case of CUDA-AWARE MPI
2172 // UCX creates CUDA buffers which are cleaned-up as part of MPI_Finalize()
2173 // resetting the device before MPI_Finalize() results in crashes inside UCX
2174 releaseDevice(deviceInfo);
2177 /* Does what it says */
2178 print_date_and_time(fplog, cr->nodeid, "Finished mdrun", gmx_gettime());
2179 walltime_accounting_destroy(walltime_accounting);
2181 // Ensure log file content is written
2184 gmx_fio_flush(logFileHandle);
2187 /* Reset FPEs (important for unit tests) by disabling them. Assumes no
2188 * exceptions were enabled before function was called. */
2191 gmx_fedisableexcept();
2194 auto rc = static_cast<int>(gmx_get_stop_condition());
2197 /* we need to join all threads. The sub-threads join when they
2198 exit this function, but the master thread needs to be told to
2208 Mdrunner::~Mdrunner()
2210 // Clean up of the Manager.
2211 // This will end up getting called on every thread-MPI rank, which is unnecessary,
2212 // but okay as long as threads synchronize some time before adding or accessing
2213 // a new set of restraints.
2214 if (restraintManager_)
2216 restraintManager_->clear();
2217 GMX_ASSERT(restraintManager_->countRestraints() == 0,
2218 "restraints added during runner life time should be cleared at runner "
2223 void Mdrunner::addPotential(std::shared_ptr<gmx::IRestraintPotential> puller, const std::string& name)
2225 GMX_ASSERT(restraintManager_, "Mdrunner must have a restraint manager.");
2226 // Not sure if this should be logged through the md logger or something else,
2227 // but it is helpful to have some sort of INFO level message sent somewhere.
2228 // std::cout << "Registering restraint named " << name << std::endl;
2230 // When multiple restraints are used, it may be wasteful to register them separately.
2231 // Maybe instead register an entire Restraint Manager as a force provider.
2232 restraintManager_->addToSpec(std::move(puller), name);
2235 Mdrunner::Mdrunner(std::unique_ptr<MDModules> mdModules) : mdModules_(std::move(mdModules)) {}
2237 Mdrunner::Mdrunner(Mdrunner&&) noexcept = default;
2239 //NOLINTNEXTLINE(performance-noexcept-move-constructor) working around GCC bug 58265 in CentOS 7
2240 Mdrunner& Mdrunner::operator=(Mdrunner&& /*handle*/) noexcept(BUGFREE_NOEXCEPT_STRING) = default;
2242 class Mdrunner::BuilderImplementation
2245 BuilderImplementation() = delete;
2246 BuilderImplementation(std::unique_ptr<MDModules> mdModules, compat::not_null<SimulationContext*> context);
2247 ~BuilderImplementation();
2249 BuilderImplementation& setExtraMdrunOptions(const MdrunOptions& options,
2250 real forceWarningThreshold,
2251 StartingBehavior startingBehavior);
2253 void addHardwareDetectionResult(const gmx_hw_info_t* hwinfo);
2255 void addDomdec(const DomdecOptions& options);
2257 void addInput(SimulationInputHandle inputHolder);
2259 void addVerletList(int nstlist);
2261 void addReplicaExchange(const ReplicaExchangeParameters& params);
2263 void addNonBonded(const char* nbpu_opt);
2265 void addPME(const char* pme_opt_, const char* pme_fft_opt_);
2267 void addBondedTaskAssignment(const char* bonded_opt);
2269 void addUpdateTaskAssignment(const char* update_opt);
2271 void addHardwareOptions(const gmx_hw_opt_t& hardwareOptions);
2273 void addFilenames(ArrayRef<const t_filenm> filenames);
2275 void addOutputEnvironment(gmx_output_env_t* outputEnvironment);
2277 void addLogFile(t_fileio* logFileHandle);
2279 void addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder);
2284 // Default parameters copied from runner.h
2285 // \todo Clarify source(s) of default parameters.
2287 const char* nbpu_opt_ = nullptr;
2288 const char* pme_opt_ = nullptr;
2289 const char* pme_fft_opt_ = nullptr;
2290 const char* bonded_opt_ = nullptr;
2291 const char* update_opt_ = nullptr;
2293 MdrunOptions mdrunOptions_;
2295 DomdecOptions domdecOptions_;
2297 ReplicaExchangeParameters replicaExchangeParameters_;
2299 //! Command-line override for the duration of a neighbor list with the Verlet scheme.
2302 //! World communicator, used for hardware detection and task assignment
2303 MPI_Comm libraryWorldCommunicator_ = MPI_COMM_NULL;
2305 //! Multisim communicator handle.
2306 gmx_multisim_t* multiSimulation_;
2308 //! mdrun communicator
2309 MPI_Comm simulationCommunicator_ = MPI_COMM_NULL;
2311 //! Print a warning if any force is larger than this (in kJ/mol nm).
2312 real forceWarningThreshold_ = -1;
2314 //! Whether the simulation will start afresh, or restart with/without appending.
2315 StartingBehavior startingBehavior_ = StartingBehavior::NewSimulation;
2317 //! The modules that comprise the functionality of mdrun.
2318 std::unique_ptr<MDModules> mdModules_;
2320 //! Detected hardware.
2321 const gmx_hw_info_t* hwinfo_ = nullptr;
2323 //! \brief Parallelism information.
2324 gmx_hw_opt_t hardwareOptions_;
2326 //! filename options for simulation.
2327 ArrayRef<const t_filenm> filenames_;
2329 /*! \brief Handle to output environment.
2331 * \todo gmx_output_env_t needs lifetime management.
2333 gmx_output_env_t* outputEnvironment_ = nullptr;
2335 /*! \brief Non-owning handle to MD log file.
2337 * \todo Context should own output facilities for client.
2338 * \todo Improve log file handle management.
2340 * Code managing the FILE* relies on the ability to set it to
2341 * nullptr to check whether the filehandle is valid.
2343 t_fileio* logFileHandle_ = nullptr;
2346 * \brief Builder for simulation stop signal handler.
2348 std::unique_ptr<StopHandlerBuilder> stopHandlerBuilder_ = nullptr;
2351 * \brief Sources for initial simulation state.
2353 * See issue #3652 for near-term refinements to the SimulationInput interface.
2355 * See issue #3379 for broader discussion on API aspects of simulation inputs and outputs.
2357 SimulationInputHandle inputHolder_;
2360 Mdrunner::BuilderImplementation::BuilderImplementation(std::unique_ptr<MDModules> mdModules,
2361 compat::not_null<SimulationContext*> context) :
2362 mdModules_(std::move(mdModules))
2364 libraryWorldCommunicator_ = context->libraryWorldCommunicator_;
2365 simulationCommunicator_ = context->simulationCommunicator_;
2366 multiSimulation_ = context->multiSimulation_.get();
2369 Mdrunner::BuilderImplementation::~BuilderImplementation() = default;
2371 Mdrunner::BuilderImplementation&
2372 Mdrunner::BuilderImplementation::setExtraMdrunOptions(const MdrunOptions& options,
2373 const real forceWarningThreshold,
2374 const StartingBehavior startingBehavior)
2376 mdrunOptions_ = options;
2377 forceWarningThreshold_ = forceWarningThreshold;
2378 startingBehavior_ = startingBehavior;
2382 void Mdrunner::BuilderImplementation::addDomdec(const DomdecOptions& options)
2384 domdecOptions_ = options;
2387 void Mdrunner::BuilderImplementation::addVerletList(int nstlist)
2392 void Mdrunner::BuilderImplementation::addReplicaExchange(const ReplicaExchangeParameters& params)
2394 replicaExchangeParameters_ = params;
2397 Mdrunner Mdrunner::BuilderImplementation::build()
2399 auto newRunner = Mdrunner(std::move(mdModules_));
2401 newRunner.mdrunOptions = mdrunOptions_;
2402 newRunner.pforce = forceWarningThreshold_;
2403 newRunner.startingBehavior = startingBehavior_;
2404 newRunner.domdecOptions = domdecOptions_;
2406 // \todo determine an invariant to check or confirm that all gmx_hw_opt_t objects are valid
2407 newRunner.hw_opt = hardwareOptions_;
2409 // No invariant to check. This parameter exists to optionally override other behavior.
2410 newRunner.nstlist_cmdline = nstlist_;
2412 newRunner.replExParams = replicaExchangeParameters_;
2414 newRunner.filenames = filenames_;
2416 newRunner.libraryWorldCommunicator = libraryWorldCommunicator_;
2418 newRunner.simulationCommunicator = simulationCommunicator_;
2420 // nullptr is a valid value for the multisim handle
2421 newRunner.ms = multiSimulation_;
2425 newRunner.hwinfo_ = hwinfo_;
2429 GMX_THROW(gmx::APIError(
2430 "MdrunnerBuilder::addHardwareDetectionResult() is required before build()"));
2435 newRunner.inputHolder_ = std::move(inputHolder_);
2439 GMX_THROW(gmx::APIError("MdrunnerBuilder::addInput() is required before build()."));
2442 // \todo Clarify ownership and lifetime management for gmx_output_env_t
2443 // \todo Update sanity checking when output environment has clearly specified invariants.
2444 // Initialization and default values for oenv are not well specified in the current version.
2445 if (outputEnvironment_)
2447 newRunner.oenv = outputEnvironment_;
2451 GMX_THROW(gmx::APIError(
2452 "MdrunnerBuilder::addOutputEnvironment() is required before build()"));
2455 newRunner.logFileHandle = logFileHandle_;
2459 newRunner.nbpu_opt = nbpu_opt_;
2463 GMX_THROW(gmx::APIError("MdrunnerBuilder::addNonBonded() is required before build()"));
2466 if (pme_opt_ && pme_fft_opt_)
2468 newRunner.pme_opt = pme_opt_;
2469 newRunner.pme_fft_opt = pme_fft_opt_;
2473 GMX_THROW(gmx::APIError("MdrunnerBuilder::addElectrostatics() is required before build()"));
2478 newRunner.bonded_opt = bonded_opt_;
2482 GMX_THROW(gmx::APIError(
2483 "MdrunnerBuilder::addBondedTaskAssignment() is required before build()"));
2488 newRunner.update_opt = update_opt_;
2492 GMX_THROW(gmx::APIError(
2493 "MdrunnerBuilder::addUpdateTaskAssignment() is required before build() "));
2497 newRunner.restraintManager_ = std::make_unique<gmx::RestraintManager>();
2499 if (stopHandlerBuilder_)
2501 newRunner.stopHandlerBuilder_ = std::move(stopHandlerBuilder_);
2505 newRunner.stopHandlerBuilder_ = std::make_unique<StopHandlerBuilder>();
2511 void Mdrunner::BuilderImplementation::addHardwareDetectionResult(const gmx_hw_info_t* hwinfo)
2516 void Mdrunner::BuilderImplementation::addNonBonded(const char* nbpu_opt)
2518 nbpu_opt_ = nbpu_opt;
2521 void Mdrunner::BuilderImplementation::addPME(const char* pme_opt, const char* pme_fft_opt)
2524 pme_fft_opt_ = pme_fft_opt;
2527 void Mdrunner::BuilderImplementation::addBondedTaskAssignment(const char* bonded_opt)
2529 bonded_opt_ = bonded_opt;
2532 void Mdrunner::BuilderImplementation::addUpdateTaskAssignment(const char* update_opt)
2534 update_opt_ = update_opt;
2537 void Mdrunner::BuilderImplementation::addHardwareOptions(const gmx_hw_opt_t& hardwareOptions)
2539 hardwareOptions_ = hardwareOptions;
2542 void Mdrunner::BuilderImplementation::addFilenames(ArrayRef<const t_filenm> filenames)
2544 filenames_ = filenames;
2547 void Mdrunner::BuilderImplementation::addOutputEnvironment(gmx_output_env_t* outputEnvironment)
2549 outputEnvironment_ = outputEnvironment;
2552 void Mdrunner::BuilderImplementation::addLogFile(t_fileio* logFileHandle)
2554 logFileHandle_ = logFileHandle;
2557 void Mdrunner::BuilderImplementation::addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder)
2559 stopHandlerBuilder_ = std::move(builder);
2562 void Mdrunner::BuilderImplementation::addInput(SimulationInputHandle inputHolder)
2564 inputHolder_ = std::move(inputHolder);
2567 MdrunnerBuilder::MdrunnerBuilder(std::unique_ptr<MDModules> mdModules,
2568 compat::not_null<SimulationContext*> context) :
2569 impl_{ std::make_unique<Mdrunner::BuilderImplementation>(std::move(mdModules), context) }
2573 MdrunnerBuilder::~MdrunnerBuilder() = default;
2575 MdrunnerBuilder& MdrunnerBuilder::addHardwareDetectionResult(const gmx_hw_info_t* hwinfo)
2577 impl_->addHardwareDetectionResult(hwinfo);
2581 MdrunnerBuilder& MdrunnerBuilder::addSimulationMethod(const MdrunOptions& options,
2582 real forceWarningThreshold,
2583 const StartingBehavior startingBehavior)
2585 impl_->setExtraMdrunOptions(options, forceWarningThreshold, startingBehavior);
2589 MdrunnerBuilder& MdrunnerBuilder::addDomainDecomposition(const DomdecOptions& options)
2591 impl_->addDomdec(options);
2595 MdrunnerBuilder& MdrunnerBuilder::addNeighborList(int nstlist)
2597 impl_->addVerletList(nstlist);
2601 MdrunnerBuilder& MdrunnerBuilder::addReplicaExchange(const ReplicaExchangeParameters& params)
2603 impl_->addReplicaExchange(params);
2607 MdrunnerBuilder& MdrunnerBuilder::addNonBonded(const char* nbpu_opt)
2609 impl_->addNonBonded(nbpu_opt);
2613 MdrunnerBuilder& MdrunnerBuilder::addElectrostatics(const char* pme_opt, const char* pme_fft_opt)
2615 // The builder method may become more general in the future, but in this version,
2616 // parameters for PME electrostatics are both required and the only parameters
2618 if (pme_opt && pme_fft_opt)
2620 impl_->addPME(pme_opt, pme_fft_opt);
2625 gmx::InvalidInputError("addElectrostatics() arguments must be non-null pointers."));
2630 MdrunnerBuilder& MdrunnerBuilder::addBondedTaskAssignment(const char* bonded_opt)
2632 impl_->addBondedTaskAssignment(bonded_opt);
2636 MdrunnerBuilder& MdrunnerBuilder::addUpdateTaskAssignment(const char* update_opt)
2638 impl_->addUpdateTaskAssignment(update_opt);
2642 Mdrunner MdrunnerBuilder::build()
2644 return impl_->build();
2647 MdrunnerBuilder& MdrunnerBuilder::addHardwareOptions(const gmx_hw_opt_t& hardwareOptions)
2649 impl_->addHardwareOptions(hardwareOptions);
2653 MdrunnerBuilder& MdrunnerBuilder::addFilenames(ArrayRef<const t_filenm> filenames)
2655 impl_->addFilenames(filenames);
2659 MdrunnerBuilder& MdrunnerBuilder::addOutputEnvironment(gmx_output_env_t* outputEnvironment)
2661 impl_->addOutputEnvironment(outputEnvironment);
2665 MdrunnerBuilder& MdrunnerBuilder::addLogFile(t_fileio* logFileHandle)
2667 impl_->addLogFile(logFileHandle);
2671 MdrunnerBuilder& MdrunnerBuilder::addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder)
2673 impl_->addStopHandlerBuilder(std::move(builder));
2677 MdrunnerBuilder& MdrunnerBuilder::addInput(SimulationInputHandle input)
2679 impl_->addInput(std::move(input));
2683 MdrunnerBuilder::MdrunnerBuilder(MdrunnerBuilder&&) noexcept = default;
2685 MdrunnerBuilder& MdrunnerBuilder::operator=(MdrunnerBuilder&&) noexcept = default;