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 = (GMX_GPU_CUDA || GMX_GPU_SYCL) && useGpuForNonbonded
207 && (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,
912 userGpuTaskAssignment,
915 hw_opt.nthreads_tmpi,
916 domdecOptions.numPmeRanks);
918 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
920 /* Determine how many thread-MPI ranks to start.
922 * TODO Over-writing the user-supplied value here does
923 * prevent any possible subsequent checks from working
925 hw_opt.nthreads_tmpi = get_nthreads_mpi(hwinfo_,
933 membedHolder.doMembed());
935 // Now start the threads for thread MPI.
936 spawnThreads(hw_opt.nthreads_tmpi);
937 // The spawned threads enter mdrunner() and execution of
938 // master and spawned threads joins at the end of this block.
941 GMX_RELEASE_ASSERT(!GMX_MPI || ms || simulationCommunicator != MPI_COMM_NULL,
942 "Must have valid communicator unless running a multi-simulation");
943 CommrecHandle crHandle = init_commrec(simulationCommunicator);
944 t_commrec* cr = crHandle.get();
945 GMX_RELEASE_ASSERT(cr != nullptr, "Must have valid commrec");
947 PhysicalNodeCommunicator physicalNodeComm(libraryWorldCommunicator, gmx_physicalnode_id_hash());
949 // If we detected the topology on this system, double-check that it makes sense
950 if (hwinfo_->hardwareTopology->isThisSystem())
952 hardwareTopologyDoubleCheckDetection(mdlog, *hwinfo_->hardwareTopology);
957 /* now broadcast everything to the non-master nodes/threads: */
958 if (!isSimulationMasterRank)
960 // Until now, only the master rank has a non-null pointer.
961 // On non-master ranks, allocate the object that will receive data in the following call.
962 inputrec = std::make_unique<t_inputrec>();
964 init_parallel(cr->mpiDefaultCommunicator,
968 partialDeserializedTpr.get());
970 GMX_RELEASE_ASSERT(inputrec != nullptr, "All ranks should have a valid inputrec now");
971 partialDeserializedTpr.reset(nullptr);
973 // Now the number of ranks is known to all ranks, and each knows
974 // the inputrec read by the master rank. The ranks can now all run
975 // the task-deciding functions and will agree on the result
976 // without needing to communicate.
977 // The LBFGS minimizer, test-particle insertion, normal modes and shell dynamics don't support DD
978 const bool useDomainDecomposition =
979 !(inputrec->eI == IntegrationAlgorithm::LBFGS || EI_TPI(inputrec->eI)
980 || inputrec->eI == IntegrationAlgorithm::NM
981 || gmx_mtop_particletype_count(mtop)[ParticleType::Shell] > 0);
983 // Note that these variables describe only their own node.
985 // Note that when bonded interactions run on a GPU they always run
986 // alongside a nonbonded task, so do not influence task assignment
987 // even though they affect the force calculation workload.
988 bool useGpuForNonbonded = false;
989 bool useGpuForPme = false;
990 bool useGpuForBonded = false;
991 bool useGpuForUpdate = false;
992 bool gpusWereDetected = hwinfo_->ngpu_compatible_tot > 0;
995 // It's possible that there are different numbers of GPUs on
996 // different nodes, which is the user's responsibility to
997 // handle. If unsuitable, we will notice that during task
999 auto canUseGpuForNonbonded = buildSupportsNonbondedOnGpu(nullptr);
1000 useGpuForNonbonded = decideWhetherToUseGpusForNonbonded(
1002 userGpuTaskAssignment,
1003 emulateGpuNonbonded,
1004 canUseGpuForNonbonded,
1005 gpuAccelerationOfNonbondedIsUseful(mdlog, *inputrec, !GMX_THREAD_MPI),
1007 useGpuForPme = decideWhetherToUseGpusForPme(useGpuForNonbonded,
1010 userGpuTaskAssignment,
1013 cr->sizeOfDefaultCommunicator,
1014 domdecOptions.numPmeRanks,
1016 useGpuForBonded = decideWhetherToUseGpusForBonded(
1017 useGpuForNonbonded, useGpuForPme, bondedTarget, *inputrec, mtop, domdecOptions.numPmeRanks, gpusWereDetected);
1019 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1021 const PmeRunMode pmeRunMode = determinePmeRunMode(useGpuForPme, pmeFftTarget, *inputrec);
1023 // Initialize development feature flags that enabled by environment variable
1024 // and report those features that are enabled.
1025 const DevelopmentFeatureFlags devFlags =
1026 manageDevelopmentFeatures(mdlog, useGpuForNonbonded, pmeRunMode);
1028 const bool useModularSimulator = checkUseModularSimulator(false,
1035 doEssentialDynamics,
1036 membedHolder.doMembed());
1038 ObservablesReducerBuilder observablesReducerBuilder;
1040 // Build restraints.
1041 // TODO: hide restraint implementation details from Mdrunner.
1042 // There is nothing unique about restraints at this point as far as the
1043 // Mdrunner is concerned. The Mdrunner should just be getting a sequence of
1044 // factory functions from the SimulationContext on which to call mdModules_->add().
1045 // TODO: capture all restraints into a single RestraintModule, passed to the runner builder.
1046 for (auto&& restraint : restraintManager_->getRestraints())
1048 auto module = RestraintMDModule::create(restraint, restraint->sites());
1049 mdModules_->add(std::move(module));
1052 // TODO: Error handling
1053 mdModules_->assignOptionsToModules(*inputrec->params, nullptr);
1054 // now that the MDModules know their options, they know which callbacks to sign up to
1055 mdModules_->subscribeToSimulationSetupNotifications();
1056 const auto& setupNotifier = mdModules_->notifiers().simulationSetupNotifier_;
1058 // Notify MdModules of existing logger
1059 setupNotifier.notify(mdlog);
1061 // Notify MdModules of internal parameters, saved into KVT
1062 if (inputrec->internalParameters != nullptr)
1064 setupNotifier.notify(*inputrec->internalParameters);
1067 // Let MdModules know the .tpr filename
1069 gmx::MdRunInputFilename mdRunInputFilename = { ftp2fn(efTPR, filenames.size(), filenames.data()) };
1070 setupNotifier.notify(mdRunInputFilename);
1073 if (fplog != nullptr)
1075 pr_inputrec(fplog, 0, "Input Parameters", inputrec.get(), FALSE);
1076 fprintf(fplog, "\n");
1081 /* In rerun, set velocities to zero if present */
1082 if (doRerun && ((globalState->flags & enumValueToBitMask(StateEntry::V)) != 0))
1084 // rerun does not use velocities
1088 "Rerun trajectory contains velocities. Rerun does only evaluate "
1089 "potential energy and forces. The velocities will be ignored.");
1090 for (int i = 0; i < globalState->natoms; i++)
1092 clear_rvec(globalState->v[i]);
1094 globalState->flags &= ~enumValueToBitMask(StateEntry::V);
1097 /* now make sure the state is initialized and propagated */
1098 set_state_entries(globalState.get(), inputrec.get(), useModularSimulator);
1101 /* NM and TPI parallelize over force/energy calculations, not atoms,
1102 * so we need to initialize and broadcast the global state.
1104 if (inputrec->eI == IntegrationAlgorithm::NM || inputrec->eI == IntegrationAlgorithm::TPI)
1108 globalState = std::make_unique<t_state>();
1110 broadcastStateWithoutDynamics(
1111 cr->mpiDefaultCommunicator, haveDDAtomOrdering(*cr), PAR(cr), globalState.get());
1114 /* A parallel command line option consistency check that we can
1115 only do after any threads have started. */
1117 && (domdecOptions.numCells[XX] > 1 || domdecOptions.numCells[YY] > 1
1118 || domdecOptions.numCells[ZZ] > 1 || domdecOptions.numPmeRanks > 0))
1121 "The -dd or -npme option request a parallel simulation, "
1123 "but %s was compiled without threads or MPI enabled",
1124 output_env_get_program_display_name(oenv));
1125 #elif GMX_THREAD_MPI
1126 "but the number of MPI-threads (option -ntmpi) is not set or is 1");
1128 "but %s was not started through mpirun/mpiexec or only one rank was requested "
1129 "through mpirun/mpiexec",
1130 output_env_get_program_display_name(oenv));
1134 if (doRerun && (EI_ENERGY_MINIMIZATION(inputrec->eI) || IntegrationAlgorithm::NM == inputrec->eI))
1137 "The .mdp file specified an energy mininization or normal mode algorithm, and "
1138 "these are not compatible with mdrun -rerun");
1141 /* NMR restraints must be initialized before load_checkpoint,
1142 * since with time averaging the history is added to t_state.
1143 * For proper consistency check we therefore need to extend
1145 * So the PME-only nodes (if present) will also initialize
1146 * the distance restraints.
1149 /* This needs to be called before read_checkpoint to extend the state */
1150 t_disresdata* disresdata;
1151 snew(disresdata, 1);
1155 DisResRunMode::MDRun,
1156 MASTER(cr) ? DDRole::Master : DDRole::Agent,
1157 PAR(cr) ? NumRanks::Multiple : NumRanks::Single,
1162 replExParams.exchangeInterval > 0);
1164 if (gmx_mtop_ftype_count(mtop, F_ORIRES) > 0 && isSimulationMasterRank)
1166 extendStateWithOriresHistory(mtop, *inputrec, globalState.get());
1169 auto deform = prepareBoxDeformation(globalState != nullptr ? globalState->box : box,
1170 MASTER(cr) ? DDRole::Master : DDRole::Agent,
1171 PAR(cr) ? NumRanks::Multiple : NumRanks::Single,
1172 cr->mpi_comm_mygroup,
1176 /* We have to remember the generation's first step before reading checkpoint.
1177 This way, we can report to the F@H core both the generation's first step
1178 and the restored first step, thus making it able to distinguish between
1179 an interruption/resume and start of the n-th generation simulation.
1180 Having this information, the F@H core can correctly calculate and report
1183 int gen_first_step = 0;
1186 gen_first_step = inputrec->init_step;
1190 ObservablesHistory observablesHistory = {};
1192 auto modularSimulatorCheckpointData = std::make_unique<ReadCheckpointDataHolder>();
1193 if (startingBehavior != StartingBehavior::NewSimulation)
1195 /* Check if checkpoint file exists before doing continuation.
1196 * This way we can use identical input options for the first and subsequent runs...
1198 if (mdrunOptions.numStepsCommandline > -2)
1200 /* Temporarily set the number of steps to unlimited to avoid
1201 * triggering the nsteps check in load_checkpoint().
1202 * This hack will go away soon when the -nsteps option is removed.
1204 inputrec->nsteps = -1;
1207 // Finish applying initial simulation state information from external sources on all ranks.
1208 // Reconcile checkpoint file data with Mdrunner state established up to this point.
1209 applyLocalState(*inputHolder_.get(),
1212 domdecOptions.numCells,
1215 &observablesHistory,
1216 mdrunOptions.reproducible,
1217 mdModules_->notifiers(),
1218 modularSimulatorCheckpointData.get(),
1219 useModularSimulator);
1220 // TODO: (#3652) Synchronize filesystem state, SimulationInput contents, and program
1222 // on all code paths.
1223 // Write checkpoint or provide hook to update SimulationInput.
1224 // If there was a checkpoint file, SimulationInput contains more information
1225 // than if there wasn't. At this point, we have synchronized the in-memory
1226 // state with the filesystem state only for restarted simulations. We should
1227 // be calling applyLocalState unconditionally and expect that the completeness
1228 // of SimulationInput is not dependent on its creation method.
1230 if (startingBehavior == StartingBehavior::RestartWithAppending && logFileHandle)
1232 // Now we can start normal logging to the truncated log file.
1233 fplog = gmx_fio_getfp(logFileHandle);
1234 prepareLogAppending(fplog);
1235 logOwner = buildLogger(fplog, MASTER(cr));
1236 mdlog = logOwner.logger();
1243 fcRegisterSteps(inputrec->nsteps + inputrec->init_step, gen_first_step);
1247 if (mdrunOptions.numStepsCommandline > -2)
1252 "The -nsteps functionality is deprecated, and may be removed in a future "
1254 "Consider using gmx convert-tpr -nsteps or changing the appropriate .mdp "
1257 /* override nsteps with value set on the commandline */
1258 override_nsteps_cmdline(mdlog, mdrunOptions.numStepsCommandline, inputrec.get());
1260 if (isSimulationMasterRank)
1262 copy_mat(globalState->box, box);
1267 gmx_bcast(sizeof(box), box, cr->mpiDefaultCommunicator);
1270 if (inputrec->cutoff_scheme != CutoffScheme::Verlet)
1273 "This group-scheme .tpr file can no longer be run by mdrun. Please update to the "
1274 "Verlet scheme, or use an earlier version of GROMACS if necessary.");
1276 /* Update rlist and nstlist. */
1277 /* Note: prepare_verlet_scheme is calling increaseNstlist(...), which (while attempting to
1278 * increase rlist) tries to check if the newly chosen value fits with the DD scheme. As this is
1279 * run before any DD scheme is set up, this check is never executed. See #3334 for more details.
1281 prepare_verlet_scheme(fplog,
1287 useGpuForNonbonded || (emulateGpuNonbonded == EmulateGpuNonbonded::Yes),
1290 // We need to decide on update groups early, as this affects
1291 // inter-domain communication distances.
1292 auto updateGroupingsPerMoleculeType = makeUpdateGroupingsPerMoleculeType(mtop);
1293 const real maxUpdateGroupRadius = computeMaxUpdateGroupRadius(
1294 mtop, updateGroupingsPerMoleculeType, maxReferenceTemperature(*inputrec));
1295 const real cutoffMargin = std::sqrt(max_cutoff2(inputrec->pbcType, box)) - inputrec->rlist;
1296 UpdateGroups updateGroups = makeUpdateGroups(mdlog,
1297 std::move(updateGroupingsPerMoleculeType),
1298 maxUpdateGroupRadius,
1299 useDomainDecomposition,
1300 systemHasConstraintsOrVsites(mtop),
1305 const bool haveFrozenAtoms = inputrecFrozenAtoms(inputrec.get());
1307 useGpuForUpdate = decideWhetherToUseGpuForUpdate(useDomainDecomposition,
1308 updateGroups.useUpdateGroups(),
1310 domdecOptions.numPmeRanks > 0,
1316 doEssentialDynamics,
1317 gmx_mtop_ftype_count(mtop, F_ORIRES) > 0,
1323 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1325 bool useGpuDirectHalo = false;
1327 if (useGpuForNonbonded)
1329 // cr->npmenodes is not yet initialized.
1330 // domdecOptions.numPmeRanks == -1 results in 0 separate PME ranks when useGpuForNonbonded is true.
1331 // Todo: remove this assumption later once auto mode has support for separate PME rank
1332 const int numPmeRanks = domdecOptions.numPmeRanks > 0 ? domdecOptions.numPmeRanks : 0;
1333 bool havePPDomainDecomposition = (cr->sizeOfDefaultCommunicator - numPmeRanks) > 1;
1334 useGpuDirectHalo = decideWhetherToUseGpuForHalo(devFlags,
1335 havePPDomainDecomposition,
1337 useModularSimulator,
1339 EI_ENERGY_MINIMIZATION(inputrec->eI));
1342 // This builder is necessary while we have multi-part construction
1343 // of DD. Before DD is constructed, we use the existence of
1344 // the builder object to indicate that further construction of DD
1346 std::unique_ptr<DomainDecompositionBuilder> ddBuilder;
1347 if (useDomainDecomposition)
1349 // P2P GPU comm + GPU update leads to case in which we enqueue async work for multiple
1350 // timesteps. DLB needs to be disabled in that case
1351 const bool directGpuCommUsedWithGpuUpdate = GMX_THREAD_MPI && useGpuDirectHalo && useGpuForUpdate;
1352 ddBuilder = std::make_unique<DomainDecompositionBuilder>(
1359 mdModules_->notifiers(),
1361 updateGroups.updateGroupingPerMoleculeType(),
1362 updateGroups.useUpdateGroups(),
1363 updateGroups.maxUpdateGroupRadius(),
1364 positionsFromStatePointer(globalState.get()),
1367 directGpuCommUsedWithGpuUpdate);
1371 /* PME, if used, is done on all nodes with 1D decomposition */
1372 cr->nnodes = cr->sizeOfDefaultCommunicator;
1373 cr->sim_nodeid = cr->rankInDefaultCommunicator;
1374 cr->nodeid = cr->rankInDefaultCommunicator;
1376 cr->duty = (DUTY_PP | DUTY_PME);
1378 if (inputrec->pbcType == PbcType::Screw)
1380 gmx_fatal(FARGS, "pbc=screw is only implemented with domain decomposition");
1384 // Produce the task assignment for this rank - done after DD is constructed
1385 GpuTaskAssignments gpuTaskAssignments = GpuTaskAssignmentsBuilder::build(
1387 userGpuTaskAssignment,
1389 simulationCommunicator,
1397 thisRankHasDuty(cr, DUTY_PP),
1398 // TODO cr->duty & DUTY_PME should imply that a PME
1399 // algorithm is active, but currently does not.
1400 EEL_PME(inputrec->coulombtype) && thisRankHasDuty(cr, DUTY_PME));
1402 // Get the device handles for the modules, nullptr when no task is assigned.
1404 DeviceInformation* deviceInfo = gpuTaskAssignments.initDevice(&deviceId);
1406 // timing enabling - TODO put this in gpu_utils (even though generally this is just option handling?)
1407 bool useTiming = true;
1411 /* WARNING: CUDA timings are incorrect with multiple streams.
1412 * This is the main reason why they are disabled by default.
1414 // TODO: Consider turning on by default when we can detect nr of streams.
1415 useTiming = (getenv("GMX_ENABLE_GPU_TIMING") != nullptr);
1417 else if (GMX_GPU_OPENCL)
1419 useTiming = (getenv("GMX_DISABLE_GPU_TIMING") == nullptr);
1422 // TODO Currently this is always built, yet DD partition code
1423 // checks if it is built before using it. Probably it should
1424 // become an MDModule that is made only when another module
1425 // requires it (e.g. pull, CompEl, density fitting), so that we
1426 // don't update the local atom sets unilaterally every step.
1427 LocalAtomSetManager atomSets;
1429 // Local state and topology are declared (and perhaps constructed)
1430 // now, because DD needs them for the LocalTopologyChecker, but
1431 // they do not contain valid data until after the first DD
1433 std::unique_ptr<t_state> localStateInstance;
1434 t_state* localState;
1435 gmx_localtop_t localTopology(mtop.ffparams);
1439 localStateInstance = std::make_unique<t_state>();
1440 localState = localStateInstance.get();
1441 // TODO Pass the GPU streams to ddBuilder to use in buffer
1442 // transfers (e.g. halo exchange)
1443 cr->dd = ddBuilder->build(&atomSets, localTopology, *localState, &observablesReducerBuilder);
1444 // The builder's job is done, so destruct it
1445 ddBuilder.reset(nullptr);
1446 // Note that local state still does not exist yet.
1450 // Without DD, the local state is merely an alias to the global state,
1451 // so we don't need to allocate anything.
1452 localState = globalState.get();
1455 // Ensure that all atoms within the same update group are in the
1456 // same periodic image. Otherwise, a simulation that did not use
1457 // update groups (e.g. a single-rank simulation) cannot always be
1458 // correctly restarted in a way that does use update groups
1459 // (e.g. a multi-rank simulation).
1460 if (isSimulationMasterRank)
1462 const bool useUpdateGroups = cr->dd ? ddUsesUpdateGroups(*cr->dd) : false;
1463 if (useUpdateGroups)
1465 putUpdateGroupAtomsInSamePeriodicImage(*cr->dd, mtop, globalState->box, globalState->x);
1469 const bool printHostName = (cr->nnodes > 1);
1470 gpuTaskAssignments.reportGpuUsage(mdlog, printHostName, useGpuForBonded, pmeRunMode, useGpuForUpdate);
1472 const bool disableNonbondedCalculation = (getenv("GMX_NO_NONBONDED") != nullptr);
1473 if (disableNonbondedCalculation)
1475 /* turn off non-bonded calculations */
1476 GMX_LOG(mdlog.warning)
1479 "Found environment variable GMX_NO_NONBONDED.\n"
1480 "Disabling nonbonded calculations.");
1483 MdrunScheduleWorkload runScheduleWork;
1485 // Also populates the simulation constant workload description.
1486 // Note: currently the default duty is DUTY_PP | DUTY_PME for all simulations, including those without PME,
1487 // so this boolean is sufficient on all ranks to determine whether separate PME ranks are used,
1488 // but this will no longer be the case if cr->duty is changed for !EEL_PME(fr->ic->eeltype).
1489 const bool haveSeparatePmeRank = (!thisRankHasDuty(cr, DUTY_PP) || !thisRankHasDuty(cr, DUTY_PME));
1490 runScheduleWork.simulationWork = createSimulationWorkload(*inputrec,
1491 disableNonbondedCalculation,
1493 havePPDomainDecomposition(cr),
1494 haveSeparatePmeRank,
1501 std::unique_ptr<DeviceStreamManager> deviceStreamManager = nullptr;
1503 if (deviceInfo != nullptr)
1505 if (runScheduleWork.simulationWork.havePpDomainDecomposition && thisRankHasDuty(cr, DUTY_PP))
1507 dd_setup_dlb_resource_sharing(cr, deviceId);
1509 deviceStreamManager = std::make_unique<DeviceStreamManager>(
1510 *deviceInfo, havePPDomainDecomposition(cr), runScheduleWork.simulationWork, useTiming);
1513 // If the user chose a task assignment, give them some hints
1514 // where appropriate.
1515 if (!userGpuTaskAssignment.empty())
1517 gpuTaskAssignments.logPerformanceHints(mdlog, numAvailableDevices);
1522 /* After possible communicator splitting in make_dd_communicators.
1523 * we can set up the intra/inter node communication.
1525 gmx_setup_nodecomm(fplog, cr);
1531 GMX_LOG(mdlog.warning)
1533 .appendTextFormatted(
1534 "This is simulation %d out of %d running as a composite GROMACS\n"
1535 "multi-simulation job. Setup for this simulation:\n",
1536 ms->simulationIndex_,
1537 ms->numSimulations_);
1539 GMX_LOG(mdlog.warning)
1540 .appendTextFormatted("Using %d MPI %s\n",
1543 cr->nnodes == 1 ? "thread" : "threads"
1545 cr->nnodes == 1 ? "process" : "processes"
1551 // If mdrun -pin auto honors any affinity setting that already
1552 // exists. If so, it is nice to provide feedback about whether
1553 // that existing affinity setting was from OpenMP or something
1554 // else, so we run this code both before and after we initialize
1555 // the OpenMP support.
1556 gmx_check_thread_affinity_set(mdlog, &hw_opt, hwinfo_->nthreads_hw_avail, FALSE);
1557 /* Check and update the number of OpenMP threads requested */
1558 checkAndUpdateRequestedNumOpenmpThreads(
1559 &hw_opt, *hwinfo_, cr, ms, physicalNodeComm.size_, pmeRunMode, mtop, *inputrec);
1561 gmx_omp_nthreads_init(mdlog,
1563 hwinfo_->nthreads_hw_avail,
1564 physicalNodeComm.size_,
1565 hw_opt.nthreads_omp,
1566 hw_opt.nthreads_omp_pme,
1567 !thisRankHasDuty(cr, DUTY_PP));
1569 const bool bEnableFPE = gmxShouldEnableFPExceptions();
1570 // FIXME - reconcile with gmx_feenableexcept() call from CommandLineModuleManager::run()
1573 gmx_feenableexcept();
1576 /* Now that we know the setup is consistent, check for efficiency */
1577 check_resource_division_efficiency(
1578 hwinfo_, gpuTaskAssignments.thisRankHasAnyGpuTask(), mdrunOptions.ntompOptionIsSet, cr, mdlog);
1580 /* getting number of PP/PME threads on this MPI / tMPI rank.
1581 PME: env variable should be read only on one node to make sure it is
1582 identical everywhere;
1584 const int numThreadsOnThisRank = thisRankHasDuty(cr, DUTY_PP)
1585 ? gmx_omp_nthreads_get(ModuleMultiThread::Nonbonded)
1586 : gmx_omp_nthreads_get(ModuleMultiThread::Pme);
1587 checkHardwareOversubscription(
1588 numThreadsOnThisRank, cr->nodeid, *hwinfo_->hardwareTopology, physicalNodeComm, mdlog);
1590 // Enable Peer access between GPUs where available
1591 // Only for DD, only master PP rank needs to perform setup, and only if thread MPI plus
1592 // any of the GPU communication features are active.
1593 if (haveDDAtomOrdering(*cr) && MASTER(cr) && thisRankHasDuty(cr, DUTY_PP) && GMX_THREAD_MPI
1594 && (runScheduleWork.simulationWork.useGpuHaloExchange
1595 || runScheduleWork.simulationWork.useGpuPmePpCommunication))
1597 setupGpuDevicePeerAccess(gpuTaskAssignments.deviceIdsAssigned(), mdlog);
1600 if (hw_opt.threadAffinity != ThreadAffinity::Off)
1602 /* Before setting affinity, check whether the affinity has changed
1603 * - which indicates that probably the OpenMP library has changed it
1604 * since we first checked).
1606 gmx_check_thread_affinity_set(mdlog, &hw_opt, hwinfo_->nthreads_hw_avail, TRUE);
1608 int numThreadsOnThisNode, intraNodeThreadOffset;
1609 analyzeThreadsOnThisNode(
1610 physicalNodeComm, numThreadsOnThisRank, &numThreadsOnThisNode, &intraNodeThreadOffset);
1612 /* Set the CPU affinity */
1613 gmx_set_thread_affinity(mdlog,
1616 *hwinfo_->hardwareTopology,
1617 numThreadsOnThisRank,
1618 numThreadsOnThisNode,
1619 intraNodeThreadOffset,
1623 if (mdrunOptions.timingOptions.resetStep > -1)
1628 "The -resetstep functionality is deprecated, and may be removed in a "
1631 std::unique_ptr<gmx_wallcycle> wcycle =
1632 wallcycle_init(fplog, mdrunOptions.timingOptions.resetStep, cr);
1636 /* Master synchronizes its value of reset_counters with all nodes
1637 * including PME only nodes */
1638 int64_t reset_counters = wcycle_get_reset_counters(wcycle.get());
1639 gmx_bcast(sizeof(reset_counters), &reset_counters, cr->mpi_comm_mysim);
1640 wcycle_set_reset_counters(wcycle.get(), reset_counters);
1643 // Membrane embedding must be initialized before we call init_forcerec()
1644 membedHolder.initializeMembed(fplog,
1651 &mdrunOptions.checkpointOptions.period);
1653 const bool thisRankHasPmeGpuTask = gpuTaskAssignments.thisRankHasPmeGpuTask();
1654 std::unique_ptr<MDAtoms> mdAtoms;
1655 std::unique_ptr<VirtualSitesHandler> vsite;
1658 if (thisRankHasDuty(cr, DUTY_PP))
1660 setupNotifier.notify(*cr);
1661 setupNotifier.notify(&atomSets);
1662 setupNotifier.notify(mtop);
1663 setupNotifier.notify(inputrec->pbcType);
1664 setupNotifier.notify(SimulationTimeStep{ inputrec->delta_t });
1665 /* Initiate forcerecord */
1666 fr = std::make_unique<t_forcerec>();
1667 fr->forceProviders = mdModules_->initForceProviders();
1668 init_forcerec(fplog,
1670 runScheduleWork.simulationWork,
1676 opt2fn("-table", filenames.size(), filenames.data()),
1677 opt2fn("-tablep", filenames.size(), filenames.data()),
1678 opt2fns("-tableb", filenames.size(), filenames.data()),
1680 // Dirty hack, for fixing disres and orires should be made mdmodules
1681 fr->fcdata->disres = disresdata;
1682 if (gmx_mtop_ftype_count(mtop, F_ORIRES) > 0)
1684 fr->fcdata->orires = std::make_unique<t_oriresdata>(
1685 fplog, mtop, *inputrec, ms, globalState.get(), &atomSets);
1688 // Save a handle to device stream manager to use elsewhere in the code
1689 // TODO: Forcerec is not a correct place to store it.
1690 fr->deviceStreamManager = deviceStreamManager.get();
1692 if (runScheduleWork.simulationWork.useGpuPmePpCommunication && !thisRankHasDuty(cr, DUTY_PME))
1695 deviceStreamManager != nullptr,
1696 "GPU device stream manager should be valid in order to use PME-PP direct "
1699 deviceStreamManager->streamIsValid(DeviceStreamType::PmePpTransfer),
1700 "GPU PP-PME stream should be valid in order to use GPU PME-PP direct "
1702 fr->pmePpCommGpu = std::make_unique<gmx::PmePpCommGpu>(
1705 &cr->dd->pmeForceReceiveBuffer,
1706 deviceStreamManager->context(),
1707 deviceStreamManager->stream(DeviceStreamType::PmePpTransfer));
1710 fr->nbv = Nbnxm::init_nb_verlet(mdlog,
1715 runScheduleWork.simulationWork.useGpuNonbonded,
1716 deviceStreamManager.get(),
1720 // TODO: Move the logic below to a GPU bonded builder
1721 if (runScheduleWork.simulationWork.useGpuBonded)
1723 GMX_RELEASE_ASSERT(deviceStreamManager != nullptr,
1724 "GPU device stream manager should be valid in order to use GPU "
1725 "version of bonded forces.");
1726 fr->listedForcesGpu = std::make_unique<ListedForcesGpu>(
1728 fr->ic->epsfac * fr->fudgeQQ,
1729 deviceStreamManager->context(),
1730 deviceStreamManager->bondedStream(havePPDomainDecomposition(cr)),
1733 fr->longRangeNonbondeds = std::make_unique<CpuPpLongRangeNonbondeds>(fr->n_tpi,
1734 fr->ic->ewaldcoeff_q,
1744 /* Initialize the mdAtoms structure.
1745 * mdAtoms is not filled with atom data,
1746 * as this can not be done now with domain decomposition.
1748 mdAtoms = makeMDAtoms(fplog, mtop, *inputrec, thisRankHasPmeGpuTask);
1749 if (globalState && thisRankHasPmeGpuTask)
1751 // The pinning of coordinates in the global state object works, because we only use
1752 // PME on GPU without DD or on a separate PME rank, and because the local state pointer
1753 // points to the global state object without DD.
1754 // FIXME: MD and EM separately set up the local state - this should happen in the same
1755 // function, which should also perform the pinning.
1756 changePinningPolicy(&globalState->x, pme_get_pinning_policy());
1759 /* Initialize the virtual site communication */
1760 vsite = makeVirtualSitesHandler(
1761 mtop, cr, fr->pbcType, updateGroups.updateGroupingPerMoleculeType());
1763 calc_shifts(box, fr->shift_vec);
1765 /* With periodic molecules the charge groups should be whole at start up
1766 * and the virtual sites should not be far from their proper positions.
1768 if (!inputrec->bContinuation && MASTER(cr)
1769 && !(inputrec->pbcType != PbcType::No && inputrec->bPeriodicMols))
1771 /* Make molecules whole at start of run */
1772 if (fr->pbcType != PbcType::No)
1774 do_pbc_first_mtop(fplog, inputrec->pbcType, box, &mtop, globalState->x.rvec_array());
1778 /* Correct initial vsite positions are required
1779 * for the initial distribution in the domain decomposition
1780 * and for the initial shell prediction.
1782 constructVirtualSitesGlobal(mtop, globalState->x);
1785 // Make the DD reverse topology, now that any vsites that are present are available
1786 if (haveDDAtomOrdering(*cr))
1788 dd_make_reverse_top(fplog, cr->dd, mtop, vsite.get(), *inputrec, domdecOptions.ddBondedChecking);
1791 if (EEL_PME(fr->ic->eeltype) || EVDW_PME(fr->ic->vdwtype))
1793 ewaldcoeff_q = fr->ic->ewaldcoeff_q;
1794 ewaldcoeff_lj = fr->ic->ewaldcoeff_lj;
1799 /* This is a PME only node */
1801 GMX_ASSERT(globalState == nullptr,
1802 "We don't need the state on a PME only rank and expect it to be unitialized");
1804 ewaldcoeff_q = calc_ewaldcoeff_q(inputrec->rcoulomb, inputrec->ewald_rtol);
1805 ewaldcoeff_lj = calc_ewaldcoeff_lj(inputrec->rvdw, inputrec->ewald_rtol_lj);
1808 gmx_pme_t* sepPmeData = nullptr;
1809 // This reference hides the fact that PME data is owned by runner on PME-only ranks and by forcerec on other ranks
1810 GMX_ASSERT(thisRankHasDuty(cr, DUTY_PP) == (fr != nullptr),
1811 "Double-checking that only PME-only ranks have no forcerec");
1812 gmx_pme_t*& pmedata = fr ? fr->pmedata : sepPmeData;
1814 // TODO should live in ewald module once its testing is improved
1816 // Later, this program could contain kernels that might be later
1817 // re-used as auto-tuning progresses, or subsequent simulations
1819 PmeGpuProgramStorage pmeGpuProgram;
1820 if (thisRankHasPmeGpuTask)
1823 (deviceStreamManager != nullptr),
1824 "GPU device stream manager should be initialized in order to use GPU for PME.");
1825 GMX_RELEASE_ASSERT((deviceInfo != nullptr),
1826 "GPU device should be initialized in order to use GPU for PME.");
1827 pmeGpuProgram = buildPmeGpuProgram(deviceStreamManager->context());
1830 /* Initiate PME if necessary,
1831 * either on all nodes or on dedicated PME nodes only. */
1832 if (EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype))
1834 if (mdAtoms && mdAtoms->mdatoms())
1836 nChargePerturbed = mdAtoms->mdatoms()->nChargePerturbed;
1837 if (EVDW_PME(inputrec->vdwtype))
1839 nTypePerturbed = mdAtoms->mdatoms()->nTypePerturbed;
1842 if (cr->npmenodes > 0)
1844 /* The PME only nodes need to know nChargePerturbed(FEP on Q) and nTypePerturbed(FEP on LJ)*/
1845 gmx_bcast(sizeof(nChargePerturbed), &nChargePerturbed, cr->mpi_comm_mysim);
1846 gmx_bcast(sizeof(nTypePerturbed), &nTypePerturbed, cr->mpi_comm_mysim);
1849 if (thisRankHasDuty(cr, DUTY_PME))
1853 // TODO: This should be in the builder.
1854 GMX_RELEASE_ASSERT(!runScheduleWork.simulationWork.useGpuPme
1855 || (deviceStreamManager != nullptr),
1856 "Device stream manager should be valid in order to use GPU "
1859 !runScheduleWork.simulationWork.useGpuPme
1860 || deviceStreamManager->streamIsValid(DeviceStreamType::Pme),
1861 "GPU PME stream should be valid in order to use GPU version of PME.");
1863 const DeviceContext* deviceContext = runScheduleWork.simulationWork.useGpuPme
1864 ? &deviceStreamManager->context()
1866 const DeviceStream* pmeStream =
1867 runScheduleWork.simulationWork.useGpuPme
1868 ? &deviceStreamManager->stream(DeviceStreamType::Pme)
1871 pmedata = gmx_pme_init(cr,
1872 getNumPmeDomains(cr->dd),
1874 nChargePerturbed != 0,
1875 nTypePerturbed != 0,
1876 mdrunOptions.reproducible,
1879 gmx_omp_nthreads_get(ModuleMultiThread::Pme),
1884 pmeGpuProgram.get(),
1887 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1892 if (EI_DYNAMICS(inputrec->eI))
1894 /* Turn on signal handling on all nodes */
1896 * (A user signal from the PME nodes (if any)
1897 * is communicated to the PP nodes.
1899 signal_handler_install();
1902 pull_t* pull_work = nullptr;
1903 if (thisRankHasDuty(cr, DUTY_PP))
1905 /* Assumes uniform use of the number of OpenMP threads */
1906 walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(ModuleMultiThread::Default));
1908 if (inputrec->bPull)
1910 /* Initialize pull code */
1911 pull_work = init_pull(fplog,
1912 inputrec->pull.get(),
1917 inputrec->fepvals->init_lambda);
1918 if (inputrec->pull->bXOutAverage || inputrec->pull->bFOutAverage)
1920 initPullHistory(pull_work, &observablesHistory);
1922 if (EI_DYNAMICS(inputrec->eI) && MASTER(cr))
1924 init_pull_output_files(pull_work, filenames.size(), filenames.data(), oenv, startingBehavior);
1928 std::unique_ptr<EnforcedRotation> enforcedRotation;
1931 /* Initialize enforced rotation code */
1932 enforcedRotation = init_rot(fplog,
1945 t_swap* swap = nullptr;
1946 if (inputrec->eSwapCoords != SwapType::No)
1948 /* Initialize ion swapping code */
1949 swap = init_swapcoords(fplog,
1951 opt2fn_master("-swap", filenames.size(), filenames.data(), cr),
1954 &observablesHistory,
1962 /* Let makeConstraints know whether we have essential dynamics constraints. */
1963 auto constr = makeConstraints(mtop,
1966 doEssentialDynamics,
1969 updateGroups.useUpdateGroups(),
1974 &observablesReducerBuilder);
1976 /* Energy terms and groups */
1977 gmx_enerdata_t enerd(mtop.groups.groups[SimulationAtomGroupType::EnergyOutput].size(),
1978 inputrec->fepvals->n_lambda);
1980 // cos acceleration is only supported by md, but older tpr
1981 // files might still combine it with other integrators
1982 GMX_RELEASE_ASSERT(inputrec->cos_accel == 0.0 || inputrec->eI == IntegrationAlgorithm::MD,
1983 "cos_acceleration is only supported by integrator=md");
1985 /* Kinetic energy data */
1986 gmx_ekindata_t ekind(inputrec->opts.ngtc,
1987 inputrec->cos_accel,
1988 gmx_omp_nthreads_get(ModuleMultiThread::Update));
1990 /* Set up interactive MD (IMD) */
1991 auto imdSession = makeImdSession(inputrec.get(),
1998 MASTER(cr) ? globalState->x : gmx::ArrayRef<gmx::RVec>(),
2002 mdrunOptions.imdOptions,
2005 if (haveDDAtomOrdering(*cr))
2007 GMX_RELEASE_ASSERT(fr, "fr was NULL while cr->duty was DUTY_PP");
2008 /* This call is not included in init_domain_decomposition
2009 * because fr->atomInfoForEachMoleculeBlock is set later.
2011 makeBondedLinks(cr->dd, mtop, fr->atomInfoForEachMoleculeBlock);
2014 if (runScheduleWork.simulationWork.useGpuBufferOps)
2016 fr->gpuForceReduction[gmx::AtomLocality::Local] = std::make_unique<gmx::GpuForceReduction>(
2017 deviceStreamManager->context(),
2018 deviceStreamManager->stream(gmx::DeviceStreamType::NonBondedLocal),
2020 fr->gpuForceReduction[gmx::AtomLocality::NonLocal] = std::make_unique<gmx::GpuForceReduction>(
2021 deviceStreamManager->context(),
2022 deviceStreamManager->stream(gmx::DeviceStreamType::NonBondedNonLocal),
2026 std::unique_ptr<gmx::StatePropagatorDataGpu> stateGpu;
2027 if (gpusWereDetected
2028 && ((runScheduleWork.simulationWork.useGpuPme && thisRankHasDuty(cr, DUTY_PME))
2029 || runScheduleWork.simulationWork.useGpuBufferOps))
2031 GpuApiCallBehavior transferKind =
2032 (inputrec->eI == IntegrationAlgorithm::MD && !doRerun && !useModularSimulator)
2033 ? GpuApiCallBehavior::Async
2034 : GpuApiCallBehavior::Sync;
2035 GMX_RELEASE_ASSERT(deviceStreamManager != nullptr,
2036 "GPU device stream manager should be initialized to use GPU.");
2037 stateGpu = std::make_unique<gmx::StatePropagatorDataGpu>(
2038 *deviceStreamManager, transferKind, pme_gpu_get_block_size(fr->pmedata), wcycle.get());
2039 fr->stateGpu = stateGpu.get();
2042 GMX_ASSERT(stopHandlerBuilder_, "Runner must provide StopHandlerBuilder to simulator.");
2043 SimulatorBuilder simulatorBuilder;
2045 simulatorBuilder.add(SimulatorStateData(
2046 globalState.get(), localState, &observablesHistory, &enerd, &ekind));
2047 simulatorBuilder.add(std::move(membedHolder));
2048 simulatorBuilder.add(std::move(stopHandlerBuilder_));
2049 simulatorBuilder.add(SimulatorConfig(mdrunOptions, startingBehavior, &runScheduleWork));
2052 simulatorBuilder.add(SimulatorEnv(fplog, cr, ms, mdlog, oenv, &observablesReducerBuilder));
2053 simulatorBuilder.add(Profiling(&nrnb, walltime_accounting, wcycle.get()));
2054 simulatorBuilder.add(ConstraintsParam(
2055 constr.get(), enforcedRotation ? enforcedRotation->getLegacyEnfrot() : nullptr, vsite.get()));
2056 // TODO: Separate `fr` to a separate add, and make the `build` handle the coupling sensibly.
2057 simulatorBuilder.add(LegacyInput(
2058 static_cast<int>(filenames.size()), filenames.data(), inputrec.get(), fr.get()));
2059 simulatorBuilder.add(ReplicaExchangeParameters(replExParams));
2060 simulatorBuilder.add(InteractiveMD(imdSession.get()));
2061 simulatorBuilder.add(SimulatorModules(mdModules_->outputProvider(), mdModules_->notifiers()));
2062 simulatorBuilder.add(CenterOfMassPulling(pull_work));
2063 // Todo move to an MDModule
2064 simulatorBuilder.add(IonSwapping(swap));
2065 simulatorBuilder.add(TopologyData(mtop, &localTopology, mdAtoms.get()));
2066 simulatorBuilder.add(BoxDeformationHandle(deform.get()));
2067 simulatorBuilder.add(std::move(modularSimulatorCheckpointData));
2069 // build and run simulator object based on user-input
2070 auto simulator = simulatorBuilder.build(useModularSimulator);
2073 if (fr->pmePpCommGpu)
2075 // destroy object since it is no longer required. (This needs to be done while the GPU context still exists.)
2076 fr->pmePpCommGpu.reset();
2079 if (inputrec->bPull)
2081 finish_pull(pull_work);
2083 finish_swapcoords(swap);
2087 GMX_RELEASE_ASSERT(pmedata, "pmedata was NULL while cr->duty was not DUTY_PP");
2089 walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(ModuleMultiThread::Pme));
2090 gmx_pmeonly(pmedata,
2094 walltime_accounting,
2097 runScheduleWork.simulationWork.useGpuPmePpCommunication,
2098 deviceStreamManager.get());
2101 wallcycle_stop(wcycle.get(), WallCycleCounter::Run);
2103 /* Finish up, write some stuff
2104 * if rerunMD, don't write last frame again
2112 walltime_accounting,
2113 fr ? fr->nbv.get() : nullptr,
2115 EI_DYNAMICS(inputrec->eI) && !isMultiSim(ms));
2118 deviceStreamManager.reset(nullptr);
2122 gmx_pme_destroy(pmedata);
2126 // FIXME: this is only here to manually unpin mdAtoms->chargeA_ and state->x,
2127 // before we destroy the GPU context(s)
2128 // Pinned buffers are associated with contexts in CUDA.
2129 // As soon as we destroy GPU contexts after mdrunner() exits, these lines should go.
2130 mdAtoms.reset(nullptr);
2131 globalState.reset(nullptr);
2132 localStateInstance.reset(nullptr);
2133 mdModules_.reset(nullptr); // destruct force providers here as they might also use the GPU
2134 fr.reset(nullptr); // destruct forcerec before gpu
2135 // TODO convert to C++ so we can get rid of these frees
2138 if (!hwinfo_->deviceInfoList.empty())
2140 /* stop the GPU profiler (only CUDA) */
2144 /* With tMPI we need to wait for all ranks to finish deallocation before
2145 * destroying the CUDA context as some tMPI ranks may be sharing
2148 * This is not a concern in OpenCL where we use one context per rank.
2150 * Note: it is safe to not call the barrier on the ranks which do not use GPU,
2151 * but it is easier and more futureproof to call it on the whole node.
2153 * Note that this function needs to be called even if GPUs are not used
2154 * in this run because the PME ranks have no knowledge of whether GPUs
2155 * are used or not, but all ranks need to enter the barrier below.
2156 * \todo Remove this physical node barrier after making sure
2157 * that it's not needed anymore (with a shared GPU run).
2161 physicalNodeComm.barrier();
2164 if (!devFlags.usingCudaAwareMpi)
2166 // Don't reset GPU in case of CUDA-AWARE MPI
2167 // UCX creates CUDA buffers which are cleaned-up as part of MPI_Finalize()
2168 // resetting the device before MPI_Finalize() results in crashes inside UCX
2169 releaseDevice(deviceInfo);
2172 /* Does what it says */
2173 print_date_and_time(fplog, cr->nodeid, "Finished mdrun", gmx_gettime());
2174 walltime_accounting_destroy(walltime_accounting);
2176 // Ensure log file content is written
2179 gmx_fio_flush(logFileHandle);
2182 /* Reset FPEs (important for unit tests) by disabling them. Assumes no
2183 * exceptions were enabled before function was called. */
2186 gmx_fedisableexcept();
2189 auto rc = static_cast<int>(gmx_get_stop_condition());
2192 /* we need to join all threads. The sub-threads join when they
2193 exit this function, but the master thread needs to be told to
2203 Mdrunner::~Mdrunner()
2205 // Clean up of the Manager.
2206 // This will end up getting called on every thread-MPI rank, which is unnecessary,
2207 // but okay as long as threads synchronize some time before adding or accessing
2208 // a new set of restraints.
2209 if (restraintManager_)
2211 restraintManager_->clear();
2212 GMX_ASSERT(restraintManager_->countRestraints() == 0,
2213 "restraints added during runner life time should be cleared at runner "
2218 void Mdrunner::addPotential(std::shared_ptr<gmx::IRestraintPotential> puller, const std::string& name)
2220 GMX_ASSERT(restraintManager_, "Mdrunner must have a restraint manager.");
2221 // Not sure if this should be logged through the md logger or something else,
2222 // but it is helpful to have some sort of INFO level message sent somewhere.
2223 // std::cout << "Registering restraint named " << name << std::endl;
2225 // When multiple restraints are used, it may be wasteful to register them separately.
2226 // Maybe instead register an entire Restraint Manager as a force provider.
2227 restraintManager_->addToSpec(std::move(puller), name);
2230 Mdrunner::Mdrunner(std::unique_ptr<MDModules> mdModules) : mdModules_(std::move(mdModules)) {}
2232 Mdrunner::Mdrunner(Mdrunner&&) noexcept = default;
2234 //NOLINTNEXTLINE(performance-noexcept-move-constructor) working around GCC bug 58265 in CentOS 7
2235 Mdrunner& Mdrunner::operator=(Mdrunner&& /*handle*/) noexcept(BUGFREE_NOEXCEPT_STRING) = default;
2237 class Mdrunner::BuilderImplementation
2240 BuilderImplementation() = delete;
2241 BuilderImplementation(std::unique_ptr<MDModules> mdModules, compat::not_null<SimulationContext*> context);
2242 ~BuilderImplementation();
2244 BuilderImplementation& setExtraMdrunOptions(const MdrunOptions& options,
2245 real forceWarningThreshold,
2246 StartingBehavior startingBehavior);
2248 void addHardwareDetectionResult(const gmx_hw_info_t* hwinfo);
2250 void addDomdec(const DomdecOptions& options);
2252 void addInput(SimulationInputHandle inputHolder);
2254 void addVerletList(int nstlist);
2256 void addReplicaExchange(const ReplicaExchangeParameters& params);
2258 void addNonBonded(const char* nbpu_opt);
2260 void addPME(const char* pme_opt_, const char* pme_fft_opt_);
2262 void addBondedTaskAssignment(const char* bonded_opt);
2264 void addUpdateTaskAssignment(const char* update_opt);
2266 void addHardwareOptions(const gmx_hw_opt_t& hardwareOptions);
2268 void addFilenames(ArrayRef<const t_filenm> filenames);
2270 void addOutputEnvironment(gmx_output_env_t* outputEnvironment);
2272 void addLogFile(t_fileio* logFileHandle);
2274 void addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder);
2279 // Default parameters copied from runner.h
2280 // \todo Clarify source(s) of default parameters.
2282 const char* nbpu_opt_ = nullptr;
2283 const char* pme_opt_ = nullptr;
2284 const char* pme_fft_opt_ = nullptr;
2285 const char* bonded_opt_ = nullptr;
2286 const char* update_opt_ = nullptr;
2288 MdrunOptions mdrunOptions_;
2290 DomdecOptions domdecOptions_;
2292 ReplicaExchangeParameters replicaExchangeParameters_;
2294 //! Command-line override for the duration of a neighbor list with the Verlet scheme.
2297 //! World communicator, used for hardware detection and task assignment
2298 MPI_Comm libraryWorldCommunicator_ = MPI_COMM_NULL;
2300 //! Multisim communicator handle.
2301 gmx_multisim_t* multiSimulation_;
2303 //! mdrun communicator
2304 MPI_Comm simulationCommunicator_ = MPI_COMM_NULL;
2306 //! Print a warning if any force is larger than this (in kJ/mol nm).
2307 real forceWarningThreshold_ = -1;
2309 //! Whether the simulation will start afresh, or restart with/without appending.
2310 StartingBehavior startingBehavior_ = StartingBehavior::NewSimulation;
2312 //! The modules that comprise the functionality of mdrun.
2313 std::unique_ptr<MDModules> mdModules_;
2315 //! Detected hardware.
2316 const gmx_hw_info_t* hwinfo_ = nullptr;
2318 //! \brief Parallelism information.
2319 gmx_hw_opt_t hardwareOptions_;
2321 //! filename options for simulation.
2322 ArrayRef<const t_filenm> filenames_;
2324 /*! \brief Handle to output environment.
2326 * \todo gmx_output_env_t needs lifetime management.
2328 gmx_output_env_t* outputEnvironment_ = nullptr;
2330 /*! \brief Non-owning handle to MD log file.
2332 * \todo Context should own output facilities for client.
2333 * \todo Improve log file handle management.
2335 * Code managing the FILE* relies on the ability to set it to
2336 * nullptr to check whether the filehandle is valid.
2338 t_fileio* logFileHandle_ = nullptr;
2341 * \brief Builder for simulation stop signal handler.
2343 std::unique_ptr<StopHandlerBuilder> stopHandlerBuilder_ = nullptr;
2346 * \brief Sources for initial simulation state.
2348 * See issue #3652 for near-term refinements to the SimulationInput interface.
2350 * See issue #3379 for broader discussion on API aspects of simulation inputs and outputs.
2352 SimulationInputHandle inputHolder_;
2355 Mdrunner::BuilderImplementation::BuilderImplementation(std::unique_ptr<MDModules> mdModules,
2356 compat::not_null<SimulationContext*> context) :
2357 mdModules_(std::move(mdModules))
2359 libraryWorldCommunicator_ = context->libraryWorldCommunicator_;
2360 simulationCommunicator_ = context->simulationCommunicator_;
2361 multiSimulation_ = context->multiSimulation_.get();
2364 Mdrunner::BuilderImplementation::~BuilderImplementation() = default;
2366 Mdrunner::BuilderImplementation&
2367 Mdrunner::BuilderImplementation::setExtraMdrunOptions(const MdrunOptions& options,
2368 const real forceWarningThreshold,
2369 const StartingBehavior startingBehavior)
2371 mdrunOptions_ = options;
2372 forceWarningThreshold_ = forceWarningThreshold;
2373 startingBehavior_ = startingBehavior;
2377 void Mdrunner::BuilderImplementation::addDomdec(const DomdecOptions& options)
2379 domdecOptions_ = options;
2382 void Mdrunner::BuilderImplementation::addVerletList(int nstlist)
2387 void Mdrunner::BuilderImplementation::addReplicaExchange(const ReplicaExchangeParameters& params)
2389 replicaExchangeParameters_ = params;
2392 Mdrunner Mdrunner::BuilderImplementation::build()
2394 auto newRunner = Mdrunner(std::move(mdModules_));
2396 newRunner.mdrunOptions = mdrunOptions_;
2397 newRunner.pforce = forceWarningThreshold_;
2398 newRunner.startingBehavior = startingBehavior_;
2399 newRunner.domdecOptions = domdecOptions_;
2401 // \todo determine an invariant to check or confirm that all gmx_hw_opt_t objects are valid
2402 newRunner.hw_opt = hardwareOptions_;
2404 // No invariant to check. This parameter exists to optionally override other behavior.
2405 newRunner.nstlist_cmdline = nstlist_;
2407 newRunner.replExParams = replicaExchangeParameters_;
2409 newRunner.filenames = filenames_;
2411 newRunner.libraryWorldCommunicator = libraryWorldCommunicator_;
2413 newRunner.simulationCommunicator = simulationCommunicator_;
2415 // nullptr is a valid value for the multisim handle
2416 newRunner.ms = multiSimulation_;
2420 newRunner.hwinfo_ = hwinfo_;
2424 GMX_THROW(gmx::APIError(
2425 "MdrunnerBuilder::addHardwareDetectionResult() is required before build()"));
2430 newRunner.inputHolder_ = std::move(inputHolder_);
2434 GMX_THROW(gmx::APIError("MdrunnerBuilder::addInput() is required before build()."));
2437 // \todo Clarify ownership and lifetime management for gmx_output_env_t
2438 // \todo Update sanity checking when output environment has clearly specified invariants.
2439 // Initialization and default values for oenv are not well specified in the current version.
2440 if (outputEnvironment_)
2442 newRunner.oenv = outputEnvironment_;
2446 GMX_THROW(gmx::APIError(
2447 "MdrunnerBuilder::addOutputEnvironment() is required before build()"));
2450 newRunner.logFileHandle = logFileHandle_;
2454 newRunner.nbpu_opt = nbpu_opt_;
2458 GMX_THROW(gmx::APIError("MdrunnerBuilder::addNonBonded() is required before build()"));
2461 if (pme_opt_ && pme_fft_opt_)
2463 newRunner.pme_opt = pme_opt_;
2464 newRunner.pme_fft_opt = pme_fft_opt_;
2468 GMX_THROW(gmx::APIError("MdrunnerBuilder::addElectrostatics() is required before build()"));
2473 newRunner.bonded_opt = bonded_opt_;
2477 GMX_THROW(gmx::APIError(
2478 "MdrunnerBuilder::addBondedTaskAssignment() is required before build()"));
2483 newRunner.update_opt = update_opt_;
2487 GMX_THROW(gmx::APIError(
2488 "MdrunnerBuilder::addUpdateTaskAssignment() is required before build() "));
2492 newRunner.restraintManager_ = std::make_unique<gmx::RestraintManager>();
2494 if (stopHandlerBuilder_)
2496 newRunner.stopHandlerBuilder_ = std::move(stopHandlerBuilder_);
2500 newRunner.stopHandlerBuilder_ = std::make_unique<StopHandlerBuilder>();
2506 void Mdrunner::BuilderImplementation::addHardwareDetectionResult(const gmx_hw_info_t* hwinfo)
2511 void Mdrunner::BuilderImplementation::addNonBonded(const char* nbpu_opt)
2513 nbpu_opt_ = nbpu_opt;
2516 void Mdrunner::BuilderImplementation::addPME(const char* pme_opt, const char* pme_fft_opt)
2519 pme_fft_opt_ = pme_fft_opt;
2522 void Mdrunner::BuilderImplementation::addBondedTaskAssignment(const char* bonded_opt)
2524 bonded_opt_ = bonded_opt;
2527 void Mdrunner::BuilderImplementation::addUpdateTaskAssignment(const char* update_opt)
2529 update_opt_ = update_opt;
2532 void Mdrunner::BuilderImplementation::addHardwareOptions(const gmx_hw_opt_t& hardwareOptions)
2534 hardwareOptions_ = hardwareOptions;
2537 void Mdrunner::BuilderImplementation::addFilenames(ArrayRef<const t_filenm> filenames)
2539 filenames_ = filenames;
2542 void Mdrunner::BuilderImplementation::addOutputEnvironment(gmx_output_env_t* outputEnvironment)
2544 outputEnvironment_ = outputEnvironment;
2547 void Mdrunner::BuilderImplementation::addLogFile(t_fileio* logFileHandle)
2549 logFileHandle_ = logFileHandle;
2552 void Mdrunner::BuilderImplementation::addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder)
2554 stopHandlerBuilder_ = std::move(builder);
2557 void Mdrunner::BuilderImplementation::addInput(SimulationInputHandle inputHolder)
2559 inputHolder_ = std::move(inputHolder);
2562 MdrunnerBuilder::MdrunnerBuilder(std::unique_ptr<MDModules> mdModules,
2563 compat::not_null<SimulationContext*> context) :
2564 impl_{ std::make_unique<Mdrunner::BuilderImplementation>(std::move(mdModules), context) }
2568 MdrunnerBuilder::~MdrunnerBuilder() = default;
2570 MdrunnerBuilder& MdrunnerBuilder::addHardwareDetectionResult(const gmx_hw_info_t* hwinfo)
2572 impl_->addHardwareDetectionResult(hwinfo);
2576 MdrunnerBuilder& MdrunnerBuilder::addSimulationMethod(const MdrunOptions& options,
2577 real forceWarningThreshold,
2578 const StartingBehavior startingBehavior)
2580 impl_->setExtraMdrunOptions(options, forceWarningThreshold, startingBehavior);
2584 MdrunnerBuilder& MdrunnerBuilder::addDomainDecomposition(const DomdecOptions& options)
2586 impl_->addDomdec(options);
2590 MdrunnerBuilder& MdrunnerBuilder::addNeighborList(int nstlist)
2592 impl_->addVerletList(nstlist);
2596 MdrunnerBuilder& MdrunnerBuilder::addReplicaExchange(const ReplicaExchangeParameters& params)
2598 impl_->addReplicaExchange(params);
2602 MdrunnerBuilder& MdrunnerBuilder::addNonBonded(const char* nbpu_opt)
2604 impl_->addNonBonded(nbpu_opt);
2608 MdrunnerBuilder& MdrunnerBuilder::addElectrostatics(const char* pme_opt, const char* pme_fft_opt)
2610 // The builder method may become more general in the future, but in this version,
2611 // parameters for PME electrostatics are both required and the only parameters
2613 if (pme_opt && pme_fft_opt)
2615 impl_->addPME(pme_opt, pme_fft_opt);
2620 gmx::InvalidInputError("addElectrostatics() arguments must be non-null pointers."));
2625 MdrunnerBuilder& MdrunnerBuilder::addBondedTaskAssignment(const char* bonded_opt)
2627 impl_->addBondedTaskAssignment(bonded_opt);
2631 MdrunnerBuilder& MdrunnerBuilder::addUpdateTaskAssignment(const char* update_opt)
2633 impl_->addUpdateTaskAssignment(update_opt);
2637 Mdrunner MdrunnerBuilder::build()
2639 return impl_->build();
2642 MdrunnerBuilder& MdrunnerBuilder::addHardwareOptions(const gmx_hw_opt_t& hardwareOptions)
2644 impl_->addHardwareOptions(hardwareOptions);
2648 MdrunnerBuilder& MdrunnerBuilder::addFilenames(ArrayRef<const t_filenm> filenames)
2650 impl_->addFilenames(filenames);
2654 MdrunnerBuilder& MdrunnerBuilder::addOutputEnvironment(gmx_output_env_t* outputEnvironment)
2656 impl_->addOutputEnvironment(outputEnvironment);
2660 MdrunnerBuilder& MdrunnerBuilder::addLogFile(t_fileio* logFileHandle)
2662 impl_->addLogFile(logFileHandle);
2666 MdrunnerBuilder& MdrunnerBuilder::addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder)
2668 impl_->addStopHandlerBuilder(std::move(builder));
2672 MdrunnerBuilder& MdrunnerBuilder::addInput(SimulationInputHandle input)
2674 impl_->addInput(std::move(input));
2678 MdrunnerBuilder::MdrunnerBuilder(MdrunnerBuilder&&) noexcept = default;
2680 MdrunnerBuilder& MdrunnerBuilder::operator=(MdrunnerBuilder&&) noexcept = default;