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/simulation_workload.h"
132 #include "gromacs/mdtypes/state.h"
133 #include "gromacs/mdtypes/state_propagator_data_gpu.h"
134 #include "gromacs/modularsimulator/modularsimulator.h"
135 #include "gromacs/nbnxm/gpu_data_mgmt.h"
136 #include "gromacs/nbnxm/nbnxm.h"
137 #include "gromacs/nbnxm/pairlist_tuning.h"
138 #include "gromacs/pbcutil/pbc.h"
139 #include "gromacs/pulling/output.h"
140 #include "gromacs/pulling/pull.h"
141 #include "gromacs/pulling/pull_rotation.h"
142 #include "gromacs/restraint/manager.h"
143 #include "gromacs/restraint/restraintmdmodule.h"
144 #include "gromacs/restraint/restraintpotential.h"
145 #include "gromacs/swap/swapcoords.h"
146 #include "gromacs/taskassignment/decidegpuusage.h"
147 #include "gromacs/taskassignment/decidesimulationworkload.h"
148 #include "gromacs/taskassignment/resourcedivision.h"
149 #include "gromacs/taskassignment/taskassignment.h"
150 #include "gromacs/taskassignment/usergpuids.h"
151 #include "gromacs/timing/gpu_timing.h"
152 #include "gromacs/timing/wallcycle.h"
153 #include "gromacs/timing/wallcyclereporting.h"
154 #include "gromacs/topology/mtop_util.h"
155 #include "gromacs/trajectory/trajectoryframe.h"
156 #include "gromacs/utility/basenetwork.h"
157 #include "gromacs/utility/cstringutil.h"
158 #include "gromacs/utility/exceptions.h"
159 #include "gromacs/utility/fatalerror.h"
160 #include "gromacs/utility/filestream.h"
161 #include "gromacs/utility/gmxassert.h"
162 #include "gromacs/utility/gmxmpi.h"
163 #include "gromacs/utility/keyvaluetree.h"
164 #include "gromacs/utility/logger.h"
165 #include "gromacs/utility/loggerbuilder.h"
166 #include "gromacs/utility/mdmodulesnotifiers.h"
167 #include "gromacs/utility/physicalnodecommunicator.h"
168 #include "gromacs/utility/pleasecite.h"
169 #include "gromacs/utility/programcontext.h"
170 #include "gromacs/utility/smalloc.h"
171 #include "gromacs/utility/stringutil.h"
172 #include "gromacs/utility/mpiinfo.h"
174 #include "isimulator.h"
175 #include "membedholder.h"
176 #include "replicaexchange.h"
177 #include "simulatorbuilder.h"
183 /*! \brief Manage any development feature flag variables encountered
185 * The use of dev features indicated by environment variables is
186 * logged in order to ensure that runs with such features enabled can
187 * be identified from their log and standard output. Any cross
188 * dependencies are also checked, and if unsatisfied, a fatal error
191 * Note that some development features overrides are applied already here:
192 * the GPU communication flags are set to false in non-tMPI and non-CUDA builds.
194 * \param[in] mdlog Logger object.
195 * \param[in] useGpuForNonbonded True if the nonbonded task is offloaded in this run.
196 * \param[in] pmeRunMode The PME run mode for this run
197 * \returns The object populated with development feature flags.
199 static DevelopmentFeatureFlags manageDevelopmentFeatures(const gmx::MDLogger& mdlog,
200 const bool useGpuForNonbonded,
201 const PmeRunMode pmeRunMode)
203 DevelopmentFeatureFlags devFlags;
205 devFlags.enableGpuBufferOps =
206 GMX_GPU_CUDA && useGpuForNonbonded && (getenv("GMX_USE_GPU_BUFFER_OPS") != nullptr);
207 devFlags.enableGpuHaloExchange = GMX_MPI && GMX_GPU_CUDA && getenv("GMX_GPU_DD_COMMS") != nullptr;
208 devFlags.forceGpuUpdateDefault = (getenv("GMX_FORCE_UPDATE_DEFAULT_GPU") != nullptr) || GMX_FAHCORE;
209 devFlags.enableGpuPmePPComm = GMX_MPI && GMX_GPU_CUDA && getenv("GMX_GPU_PME_PP_COMMS") != nullptr;
211 // Direct GPU comm path is being used with CUDA_AWARE_MPI
212 // make sure underlying MPI implementation is CUDA-aware
213 if (!GMX_THREAD_MPI && (devFlags.enableGpuPmePPComm || devFlags.enableGpuHaloExchange))
215 const bool haveDetectedCudaAwareMpi =
216 (checkMpiCudaAwareSupport() == CudaAwareMpiStatus::Supported);
217 const bool forceCudaAwareMpi = (getenv("GMX_FORCE_CUDA_AWARE_MPI") != nullptr);
219 if (!haveDetectedCudaAwareMpi && forceCudaAwareMpi)
221 // CUDA-aware support not detected in MPI library but, user has forced it's use
222 GMX_LOG(mdlog.warning)
224 .appendTextFormatted(
225 "This run has forced use of 'CUDA-aware MPI'. "
226 "But, GROMACS cannot determine if underlying MPI "
227 "is CUDA-aware. GROMACS recommends use of latest openMPI version "
228 "for CUDA-aware support. "
229 "If you observe failures at runtime, try unsetting "
230 "GMX_FORCE_CUDA_AWARE_MPI environment variable.");
233 if (haveDetectedCudaAwareMpi || forceCudaAwareMpi)
235 devFlags.usingCudaAwareMpi = true;
236 GMX_LOG(mdlog.warning)
238 .appendTextFormatted(
239 "Using CUDA-aware MPI for 'GPU halo exchange' or 'GPU PME-PP "
240 "communications' feature.");
244 if (devFlags.enableGpuHaloExchange)
246 GMX_LOG(mdlog.warning)
248 .appendTextFormatted(
249 "GMX_GPU_DD_COMMS environment variable detected, but the 'GPU "
250 "halo exchange' feature will not be enabled as GROMACS couldn't "
251 "detect CUDA_aware support in underlying MPI implementation.");
252 devFlags.enableGpuHaloExchange = false;
254 if (devFlags.enableGpuPmePPComm)
256 GMX_LOG(mdlog.warning)
259 "GMX_GPU_PME_PP_COMMS environment variable detected, but the "
260 "'GPU PME-PP communications' feature will not be enabled as "
262 "detect CUDA_aware support in underlying MPI implementation.");
263 devFlags.enableGpuPmePPComm = false;
266 GMX_LOG(mdlog.warning)
268 .appendTextFormatted(
269 "GROMACS recommends use of latest OpenMPI version for CUDA-aware "
271 "If you are certain about CUDA-aware support in your MPI library, "
272 "you can force it's use by setting environment variable "
273 " GMX_FORCE_CUDA_AWARE_MPI.");
277 if (devFlags.enableGpuBufferOps)
279 GMX_LOG(mdlog.warning)
281 .appendTextFormatted(
282 "This run uses the 'GPU buffer ops' feature, enabled by the "
283 "GMX_USE_GPU_BUFFER_OPS environment variable.");
286 if (devFlags.forceGpuUpdateDefault)
288 GMX_LOG(mdlog.warning)
290 .appendTextFormatted(
291 "This run will default to '-update gpu' as requested by the "
292 "GMX_FORCE_UPDATE_DEFAULT_GPU environment variable. GPU update with domain "
293 "decomposition lacks substantial testing and should be used with caution.");
296 if (devFlags.enableGpuHaloExchange)
298 if (useGpuForNonbonded)
300 if (!devFlags.enableGpuBufferOps)
302 GMX_LOG(mdlog.warning)
304 .appendTextFormatted(
305 "Enabling GPU buffer operations required by GMX_GPU_DD_COMMS "
306 "(equivalent with GMX_USE_GPU_BUFFER_OPS=1).");
307 devFlags.enableGpuBufferOps = true;
309 GMX_LOG(mdlog.warning)
311 .appendTextFormatted(
312 "This run has requested the 'GPU halo exchange' feature, enabled by "
314 "GMX_GPU_DD_COMMS environment variable.");
318 GMX_LOG(mdlog.warning)
320 .appendTextFormatted(
321 "GMX_GPU_DD_COMMS environment variable detected, but the 'GPU "
322 "halo exchange' feature will not be enabled as nonbonded interactions "
323 "are not offloaded.");
324 devFlags.enableGpuHaloExchange = false;
328 if (devFlags.enableGpuPmePPComm)
330 if (pmeRunMode == PmeRunMode::GPU)
332 if (!devFlags.enableGpuBufferOps)
334 GMX_LOG(mdlog.warning)
336 .appendTextFormatted(
337 "Enabling GPU buffer operations required by GMX_GPU_PME_PP_COMMS "
338 "(equivalent with GMX_USE_GPU_BUFFER_OPS=1).");
339 devFlags.enableGpuBufferOps = true;
341 GMX_LOG(mdlog.warning)
343 .appendTextFormatted(
344 "This run uses the 'GPU PME-PP communications' feature, enabled "
345 "by the GMX_GPU_PME_PP_COMMS environment variable.");
349 std::string clarification;
350 if (pmeRunMode == PmeRunMode::Mixed)
353 "PME FFT and gather are not offloaded to the GPU (PME is running in mixed "
358 clarification = "PME is not offloaded to the GPU.";
360 GMX_LOG(mdlog.warning)
363 "GMX_GPU_PME_PP_COMMS environment variable detected, but the "
364 "'GPU PME-PP communications' feature was not enabled as "
366 devFlags.enableGpuPmePPComm = false;
373 /*! \brief Barrier for safe simultaneous thread access to mdrunner data
375 * Used to ensure that the master thread does not modify mdrunner during copy
376 * on the spawned threads. */
377 static void threadMpiMdrunnerAccessBarrier()
380 MPI_Barrier(MPI_COMM_WORLD);
384 Mdrunner Mdrunner::cloneOnSpawnedThread() const
386 auto newRunner = Mdrunner(std::make_unique<MDModules>());
388 // All runners in the same process share a restraint manager resource because it is
389 // part of the interface to the client code, which is associated only with the
390 // original thread. Handles to the same resources can be obtained by copy.
392 newRunner.restraintManager_ = std::make_unique<RestraintManager>(*restraintManager_);
395 // Copy members of master runner.
396 // \todo Replace with builder when Simulation context and/or runner phases are better defined.
397 // Ref https://gitlab.com/gromacs/gromacs/-/issues/2587 and https://gitlab.com/gromacs/gromacs/-/issues/2375
398 newRunner.hw_opt = hw_opt;
399 newRunner.filenames = filenames;
401 newRunner.hwinfo_ = hwinfo_;
402 newRunner.oenv = oenv;
403 newRunner.mdrunOptions = mdrunOptions;
404 newRunner.domdecOptions = domdecOptions;
405 newRunner.nbpu_opt = nbpu_opt;
406 newRunner.pme_opt = pme_opt;
407 newRunner.pme_fft_opt = pme_fft_opt;
408 newRunner.bonded_opt = bonded_opt;
409 newRunner.update_opt = update_opt;
410 newRunner.nstlist_cmdline = nstlist_cmdline;
411 newRunner.replExParams = replExParams;
412 newRunner.pforce = pforce;
413 // Give the spawned thread the newly created valid communicator
414 // for the simulation.
415 newRunner.libraryWorldCommunicator = MPI_COMM_WORLD;
416 newRunner.simulationCommunicator = MPI_COMM_WORLD;
418 newRunner.startingBehavior = startingBehavior;
419 newRunner.stopHandlerBuilder_ = std::make_unique<StopHandlerBuilder>(*stopHandlerBuilder_);
420 newRunner.inputHolder_ = inputHolder_;
422 threadMpiMdrunnerAccessBarrier();
427 /*! \brief The callback used for running on spawned threads.
429 * Obtains the pointer to the master mdrunner object from the one
430 * argument permitted to the thread-launch API call, copies it to make
431 * a new runner for this thread, reinitializes necessary data, and
432 * proceeds to the simulation. */
433 static void mdrunner_start_fn(const void* arg)
437 const auto* masterMdrunner = reinterpret_cast<const gmx::Mdrunner*>(arg);
438 /* copy the arg list to make sure that it's thread-local. This
439 doesn't copy pointed-to items, of course; fnm, cr and fplog
440 are reset in the call below, all others should be const. */
441 gmx::Mdrunner mdrunner = masterMdrunner->cloneOnSpawnedThread();
444 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
448 void Mdrunner::spawnThreads(int numThreadsToLaunch)
451 /* now spawn new threads that start mdrunner_start_fn(), while
452 the main thread returns. Thread affinity is handled later. */
453 if (tMPI_Init_fn(TRUE, numThreadsToLaunch, TMPI_AFFINITY_NONE, mdrunner_start_fn, static_cast<const void*>(this))
456 GMX_THROW(gmx::InternalError("Failed to spawn thread-MPI threads"));
459 // Give the master thread the newly created valid communicator for
461 libraryWorldCommunicator = MPI_COMM_WORLD;
462 simulationCommunicator = MPI_COMM_WORLD;
463 threadMpiMdrunnerAccessBarrier();
465 GMX_UNUSED_VALUE(numThreadsToLaunch);
466 GMX_UNUSED_VALUE(mdrunner_start_fn);
472 /*! \brief Initialize variables for Verlet scheme simulation */
473 static void prepare_verlet_scheme(FILE* fplog,
477 const gmx_mtop_t& mtop,
479 bool makeGpuPairList,
480 const gmx::CpuInfo& cpuinfo)
482 // We checked the cut-offs in grompp, but double-check here.
483 // We have PME+LJcutoff kernels for rcoulomb>rvdw.
484 if (EEL_PME_EWALD(ir->coulombtype) && ir->vdwtype == VanDerWaalsType::Cut)
486 GMX_RELEASE_ASSERT(ir->rcoulomb >= ir->rvdw,
487 "With Verlet lists and PME we should have rcoulomb>=rvdw");
491 GMX_RELEASE_ASSERT(ir->rcoulomb == ir->rvdw,
492 "With Verlet lists and no PME rcoulomb and rvdw should be identical");
494 /* For NVE simulations, we will retain the initial list buffer */
495 if (EI_DYNAMICS(ir->eI) && ir->verletbuf_tol > 0
496 && !(EI_MD(ir->eI) && ir->etc == TemperatureCoupling::No))
498 /* Update the Verlet buffer size for the current run setup */
500 /* Here we assume SIMD-enabled kernels are being used. But as currently
501 * calc_verlet_buffer_size gives the same results for 4x8 and 4x4
502 * and 4x2 gives a larger buffer than 4x4, this is ok.
504 ListSetupType listType =
505 (makeGpuPairList ? ListSetupType::Gpu : ListSetupType::CpuSimdWhenSupported);
506 VerletbufListSetup listSetup = verletbufGetSafeListSetup(listType);
508 const real rlist_new =
509 calcVerletBufferSize(mtop, det(box), *ir, ir->nstlist, ir->nstlist - 1, -1, listSetup);
511 if (rlist_new != ir->rlist)
513 if (fplog != nullptr)
516 "\nChanging rlist from %g to %g for non-bonded %dx%d atom kernels\n\n",
519 listSetup.cluster_size_i,
520 listSetup.cluster_size_j);
522 ir->rlist = rlist_new;
526 if (nstlist_cmdline > 0 && (!EI_DYNAMICS(ir->eI) || ir->verletbuf_tol <= 0))
529 "Can not set nstlist without %s",
530 !EI_DYNAMICS(ir->eI) ? "dynamics" : "verlet-buffer-tolerance");
533 if (EI_DYNAMICS(ir->eI))
535 /* Set or try nstlist values */
536 increaseNstlist(fplog, cr, ir, nstlist_cmdline, &mtop, box, makeGpuPairList, cpuinfo);
540 /*! \brief Override the nslist value in inputrec
542 * with value passed on the command line (if any)
544 static void override_nsteps_cmdline(const gmx::MDLogger& mdlog, int64_t nsteps_cmdline, t_inputrec* ir)
548 /* override with anything else than the default -2 */
549 if (nsteps_cmdline > -2)
551 char sbuf_steps[STEPSTRSIZE];
552 char sbuf_msg[STRLEN];
554 ir->nsteps = nsteps_cmdline;
555 if (EI_DYNAMICS(ir->eI) && nsteps_cmdline != -1)
558 "Overriding nsteps with value passed on the command line: %s steps, %.3g ps",
559 gmx_step_str(nsteps_cmdline, sbuf_steps),
560 fabs(nsteps_cmdline * ir->delta_t));
565 "Overriding nsteps with value passed on the command line: %s steps",
566 gmx_step_str(nsteps_cmdline, sbuf_steps));
569 GMX_LOG(mdlog.warning).asParagraph().appendText(sbuf_msg);
571 else if (nsteps_cmdline < -2)
573 gmx_fatal(FARGS, "Invalid nsteps value passed on the command line: %" PRId64, nsteps_cmdline);
575 /* Do nothing if nsteps_cmdline == -2 */
581 /*! \brief Return whether GPU acceleration of nonbondeds is supported with the given settings.
583 * If not, and if a warning may be issued, logs a warning about
584 * falling back to CPU code. With thread-MPI, only the first
585 * call to this function should have \c issueWarning true. */
586 static bool gpuAccelerationOfNonbondedIsUseful(const MDLogger& mdlog, const t_inputrec& ir, bool issueWarning)
588 bool gpuIsUseful = true;
591 if (ir.opts.ngener - ir.nwall > 1)
593 /* The GPU code does not support more than one energy group.
594 * If the user requested GPUs explicitly, a fatal error is given later.
598 "Multiple energy groups is not implemented for GPUs, falling back to the CPU. "
599 "For better performance, run on the GPU without energy groups and then do "
600 "gmx mdrun -rerun option on the trajectory with an energy group .tpr file.";
606 warning = "TPI is not implemented for GPUs.";
609 if (!gpuIsUseful && issueWarning)
611 GMX_LOG(mdlog.warning).asParagraph().appendText(warning);
617 //! Initializes the logger for mdrun.
618 static gmx::LoggerOwner buildLogger(FILE* fplog, const bool isSimulationMasterRank)
620 gmx::LoggerBuilder builder;
621 if (fplog != nullptr)
623 builder.addTargetFile(gmx::MDLogger::LogLevel::Info, fplog);
625 if (isSimulationMasterRank)
627 builder.addTargetStream(gmx::MDLogger::LogLevel::Warning, &gmx::TextOutputFile::standardError());
629 return builder.build();
632 //! Make a TaskTarget from an mdrun argument string.
633 static TaskTarget findTaskTarget(const char* optionString)
635 TaskTarget returnValue = TaskTarget::Auto;
637 if (strncmp(optionString, "auto", 3) == 0)
639 returnValue = TaskTarget::Auto;
641 else if (strncmp(optionString, "cpu", 3) == 0)
643 returnValue = TaskTarget::Cpu;
645 else if (strncmp(optionString, "gpu", 3) == 0)
647 returnValue = TaskTarget::Gpu;
651 GMX_ASSERT(false, "Option string should have been checked for sanity already");
657 //! Finish run, aggregate data to print performance info.
658 static void finish_run(FILE* fplog,
659 const gmx::MDLogger& mdlog,
661 const t_inputrec& inputrec,
663 gmx_wallcycle* wcycle,
664 gmx_walltime_accounting_t walltime_accounting,
665 nonbonded_verlet_t* nbv,
666 const gmx_pme_t* pme,
670 double nbfs = 0, mflop = 0;
671 double elapsed_time, elapsed_time_over_all_ranks, elapsed_time_over_all_threads,
672 elapsed_time_over_all_threads_over_all_ranks;
673 /* Control whether it is valid to print a report. Only the
674 simulation master may print, but it should not do so if the run
675 terminated e.g. before a scheduled reset step. This is
676 complicated by the fact that PME ranks are unaware of the
677 reason why they were sent a pmerecvqxFINISH. To avoid
678 communication deadlocks, we always do the communication for the
679 report, even if we've decided not to write the report, because
680 how long it takes to finish the run is not important when we've
681 decided not to report on the simulation performance.
683 Further, we only report performance for dynamical integrators,
684 because those are the only ones for which we plan to
685 consider doing any optimizations. */
686 bool printReport = EI_DYNAMICS(inputrec.eI) && SIMMASTER(cr);
688 if (printReport && !walltime_accounting_get_valid_finish(walltime_accounting))
690 GMX_LOG(mdlog.warning)
692 .appendText("Simulation ended prematurely, no performance report will be written.");
697 std::unique_ptr<t_nrnb> nrnbTotalStorage;
700 nrnbTotalStorage = std::make_unique<t_nrnb>();
701 nrnb_tot = nrnbTotalStorage.get();
703 MPI_Allreduce(nrnb->n.data(), nrnb_tot->n.data(), eNRNB, MPI_DOUBLE, MPI_SUM, cr->mpi_comm_mysim);
711 elapsed_time = walltime_accounting_get_time_since_reset(walltime_accounting);
712 elapsed_time_over_all_threads =
713 walltime_accounting_get_time_since_reset_over_all_threads(walltime_accounting);
717 /* reduce elapsed_time over all MPI ranks in the current simulation */
718 MPI_Allreduce(&elapsed_time, &elapsed_time_over_all_ranks, 1, MPI_DOUBLE, MPI_SUM, cr->mpi_comm_mysim);
719 elapsed_time_over_all_ranks /= cr->nnodes;
720 /* Reduce elapsed_time_over_all_threads over all MPI ranks in the
721 * current simulation. */
722 MPI_Allreduce(&elapsed_time_over_all_threads,
723 &elapsed_time_over_all_threads_over_all_ranks,
732 elapsed_time_over_all_ranks = elapsed_time;
733 elapsed_time_over_all_threads_over_all_ranks = elapsed_time_over_all_threads;
738 print_flop(fplog, nrnb_tot, &nbfs, &mflop);
741 if (thisRankHasDuty(cr, DUTY_PP) && DOMAINDECOMP(cr))
743 print_dd_statistics(cr, inputrec, fplog);
746 /* TODO Move the responsibility for any scaling by thread counts
747 * to the code that handled the thread region, so that there's a
748 * mechanism to keep cycle counting working during the transition
749 * to task parallelism. */
750 int nthreads_pp = gmx_omp_nthreads_get(ModuleMultiThread::Nonbonded);
751 int nthreads_pme = gmx_omp_nthreads_get(ModuleMultiThread::Pme);
752 wallcycle_scale_by_num_threads(
753 wcycle, thisRankHasDuty(cr, DUTY_PME) && !thisRankHasDuty(cr, DUTY_PP), nthreads_pp, nthreads_pme);
754 auto cycle_sum(wallcycle_sum(cr, wcycle));
758 auto* nbnxn_gpu_timings =
759 (nbv != nullptr && nbv->useGpu()) ? Nbnxm::gpu_get_timings(nbv->gpu_nbv) : nullptr;
760 gmx_wallclock_gpu_pme_t pme_gpu_timings = {};
762 if (pme_gpu_task_enabled(pme))
764 pme_gpu_get_timings(pme, &pme_gpu_timings);
766 wallcycle_print(fplog,
772 elapsed_time_over_all_ranks,
778 if (EI_DYNAMICS(inputrec.eI))
780 delta_t = inputrec.delta_t;
786 elapsed_time_over_all_threads_over_all_ranks,
787 elapsed_time_over_all_ranks,
788 walltime_accounting_get_nsteps_done_since_reset(walltime_accounting),
796 elapsed_time_over_all_threads_over_all_ranks,
797 elapsed_time_over_all_ranks,
798 walltime_accounting_get_nsteps_done_since_reset(walltime_accounting),
806 int Mdrunner::mdrunner()
809 std::unique_ptr<t_forcerec> fr;
810 real ewaldcoeff_q = 0;
811 real ewaldcoeff_lj = 0;
812 int nChargePerturbed = -1, nTypePerturbed = 0;
813 gmx_walltime_accounting_t walltime_accounting = nullptr;
814 MembedHolder membedHolder(filenames.size(), filenames.data());
816 /* CAUTION: threads may be started later on in this function, so
817 cr doesn't reflect the final parallel state right now */
820 /* TODO: inputrec should tell us whether we use an algorithm, not a file option */
821 const bool doEssentialDynamics = opt2bSet("-ei", filenames.size(), filenames.data());
822 const bool doRerun = mdrunOptions.rerun;
824 // Handle task-assignment related user options.
825 EmulateGpuNonbonded emulateGpuNonbonded =
826 (getenv("GMX_EMULATE_GPU") != nullptr ? EmulateGpuNonbonded::Yes : EmulateGpuNonbonded::No);
828 std::vector<int> userGpuTaskAssignment;
831 userGpuTaskAssignment = parseUserTaskAssignmentString(hw_opt.userGpuTaskAssignment);
833 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
834 auto nonbondedTarget = findTaskTarget(nbpu_opt);
835 auto pmeTarget = findTaskTarget(pme_opt);
836 auto pmeFftTarget = findTaskTarget(pme_fft_opt);
837 auto bondedTarget = findTaskTarget(bonded_opt);
838 auto updateTarget = findTaskTarget(update_opt);
840 FILE* fplog = nullptr;
841 // If we are appending, we don't write log output because we need
842 // to check that the old log file matches what the checkpoint file
843 // expects. Otherwise, we should start to write log output now if
844 // there is a file ready for it.
845 if (logFileHandle != nullptr && startingBehavior != StartingBehavior::RestartWithAppending)
847 fplog = gmx_fio_getfp(logFileHandle);
849 const bool isSimulationMasterRank = findIsSimulationMasterRank(ms, simulationCommunicator);
850 gmx::LoggerOwner logOwner(buildLogger(fplog, isSimulationMasterRank));
851 gmx::MDLogger mdlog(logOwner.logger());
853 gmx_print_detected_hardware(fplog, isSimulationMasterRank && isMasterSim(ms), mdlog, hwinfo_);
855 std::vector<int> availableDevices =
856 makeListOfAvailableDevices(hwinfo_->deviceInfoList, hw_opt.devicesSelectedByUser);
857 const int numAvailableDevices = gmx::ssize(availableDevices);
859 // Print citation requests after all software/hardware printing
860 pleaseCiteGromacs(fplog);
862 // Note: legacy program logic relies on checking whether these pointers are assigned.
863 // Objects may or may not be allocated later.
864 std::unique_ptr<t_inputrec> inputrec;
865 std::unique_ptr<t_state> globalState;
867 auto partialDeserializedTpr = std::make_unique<PartialDeserializedTprFile>();
869 if (isSimulationMasterRank)
871 // Allocate objects to be initialized by later function calls.
872 /* Only the master rank has the global state */
873 globalState = std::make_unique<t_state>();
874 inputrec = std::make_unique<t_inputrec>();
876 /* Read (nearly) all data required for the simulation
877 * and keep the partly serialized tpr contents to send to other ranks later
879 applyGlobalSimulationState(
880 *inputHolder_.get(), partialDeserializedTpr.get(), globalState.get(), inputrec.get(), &mtop);
883 /* Check and update the hardware options for internal consistency */
884 checkAndUpdateHardwareOptions(
885 mdlog, &hw_opt, isSimulationMasterRank, domdecOptions.numPmeRanks, inputrec.get());
887 if (GMX_THREAD_MPI && isSimulationMasterRank)
889 bool useGpuForNonbonded = false;
890 bool useGpuForPme = false;
893 GMX_RELEASE_ASSERT(inputrec != nullptr, "Keep the compiler happy");
895 // If the user specified the number of ranks, then we must
896 // respect that, but in default mode, we need to allow for
897 // the number of GPUs to choose the number of ranks.
898 auto canUseGpuForNonbonded = buildSupportsNonbondedOnGpu(nullptr);
899 useGpuForNonbonded = decideWhetherToUseGpusForNonbondedWithThreadMpi(
901 numAvailableDevices > 0,
902 userGpuTaskAssignment,
904 canUseGpuForNonbonded,
905 gpuAccelerationOfNonbondedIsUseful(mdlog, *inputrec, GMX_THREAD_MPI),
906 hw_opt.nthreads_tmpi);
907 useGpuForPme = decideWhetherToUseGpusForPmeWithThreadMpi(useGpuForNonbonded,
910 userGpuTaskAssignment,
913 hw_opt.nthreads_tmpi,
914 domdecOptions.numPmeRanks);
916 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
918 /* Determine how many thread-MPI ranks to start.
920 * TODO Over-writing the user-supplied value here does
921 * prevent any possible subsequent checks from working
923 hw_opt.nthreads_tmpi = get_nthreads_mpi(hwinfo_,
931 membedHolder.doMembed());
933 // Now start the threads for thread MPI.
934 spawnThreads(hw_opt.nthreads_tmpi);
935 // The spawned threads enter mdrunner() and execution of
936 // master and spawned threads joins at the end of this block.
939 GMX_RELEASE_ASSERT(ms || simulationCommunicator != MPI_COMM_NULL,
940 "Must have valid communicator unless running a multi-simulation");
941 CommrecHandle crHandle = init_commrec(simulationCommunicator);
942 t_commrec* cr = crHandle.get();
943 GMX_RELEASE_ASSERT(cr != nullptr, "Must have valid commrec");
945 PhysicalNodeCommunicator physicalNodeComm(libraryWorldCommunicator, gmx_physicalnode_id_hash());
947 // If we detected the topology on this system, double-check that it makes sense
948 if (hwinfo_->hardwareTopology->isThisSystem())
950 hardwareTopologyDoubleCheckDetection(mdlog, *hwinfo_->hardwareTopology);
955 /* now broadcast everything to the non-master nodes/threads: */
956 if (!isSimulationMasterRank)
958 // Until now, only the master rank has a non-null pointer.
959 // On non-master ranks, allocate the object that will receive data in the following call.
960 inputrec = std::make_unique<t_inputrec>();
962 init_parallel(cr->mpiDefaultCommunicator,
966 partialDeserializedTpr.get());
968 GMX_RELEASE_ASSERT(inputrec != nullptr, "All ranks should have a valid inputrec now");
969 partialDeserializedTpr.reset(nullptr);
972 !inputrec->useConstantAcceleration,
973 "Linear acceleration has been removed in GROMACS 2022, and was broken for many years "
974 "before that. Use GROMACS 4.5 or earlier if you need this feature.");
976 // Now the number of ranks is known to all ranks, and each knows
977 // the inputrec read by the master rank. The ranks can now all run
978 // the task-deciding functions and will agree on the result
979 // without needing to communicate.
980 const bool useDomainDecomposition =
981 (PAR(cr) && !(EI_TPI(inputrec->eI) || inputrec->eI == IntegrationAlgorithm::NM));
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,
1009 userGpuTaskAssignment,
1012 cr->sizeOfDefaultCommunicator,
1013 domdecOptions.numPmeRanks,
1015 useGpuForBonded = decideWhetherToUseGpusForBonded(
1016 useGpuForNonbonded, useGpuForPme, bondedTarget, *inputrec, mtop, domdecOptions.numPmeRanks, gpusWereDetected);
1018 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1020 const PmeRunMode pmeRunMode = determinePmeRunMode(useGpuForPme, pmeFftTarget, *inputrec);
1022 // Initialize development feature flags that enabled by environment variable
1023 // and report those features that are enabled.
1024 const DevelopmentFeatureFlags devFlags =
1025 manageDevelopmentFeatures(mdlog, useGpuForNonbonded, pmeRunMode);
1027 const bool useModularSimulator = checkUseModularSimulator(false,
1034 doEssentialDynamics,
1035 membedHolder.doMembed());
1037 // Build restraints.
1038 // TODO: hide restraint implementation details from Mdrunner.
1039 // There is nothing unique about restraints at this point as far as the
1040 // Mdrunner is concerned. The Mdrunner should just be getting a sequence of
1041 // factory functions from the SimulationContext on which to call mdModules_->add().
1042 // TODO: capture all restraints into a single RestraintModule, passed to the runner builder.
1043 for (auto&& restraint : restraintManager_->getRestraints())
1045 auto module = RestraintMDModule::create(restraint, restraint->sites());
1046 mdModules_->add(std::move(module));
1049 // TODO: Error handling
1050 mdModules_->assignOptionsToModules(*inputrec->params, nullptr);
1051 // now that the MDModules know their options, they know which callbacks to sign up to
1052 mdModules_->subscribeToSimulationSetupNotifications();
1053 const auto& setupNotifier = mdModules_->notifiers().simulationSetupNotifier_;
1055 // Notify MdModules of existing logger
1056 setupNotifier.notify(mdlog);
1058 // Notify MdModules of internal parameters, saved into KVT
1059 if (inputrec->internalParameters != nullptr)
1061 setupNotifier.notify(*inputrec->internalParameters);
1064 // Let MdModules know the .tpr filename
1066 gmx::MdRunInputFilename mdRunInputFilename = { ftp2fn(efTPR, filenames.size(), filenames.data()) };
1067 setupNotifier.notify(mdRunInputFilename);
1070 if (fplog != nullptr)
1072 pr_inputrec(fplog, 0, "Input Parameters", inputrec.get(), FALSE);
1073 fprintf(fplog, "\n");
1078 /* In rerun, set velocities to zero if present */
1079 if (doRerun && ((globalState->flags & enumValueToBitMask(StateEntry::V)) != 0))
1081 // rerun does not use velocities
1085 "Rerun trajectory contains velocities. Rerun does only evaluate "
1086 "potential energy and forces. The velocities will be ignored.");
1087 for (int i = 0; i < globalState->natoms; i++)
1089 clear_rvec(globalState->v[i]);
1091 globalState->flags &= ~enumValueToBitMask(StateEntry::V);
1094 /* now make sure the state is initialized and propagated */
1095 set_state_entries(globalState.get(), inputrec.get(), useModularSimulator);
1098 /* NM and TPI parallelize over force/energy calculations, not atoms,
1099 * so we need to initialize and broadcast the global state.
1101 if (inputrec->eI == IntegrationAlgorithm::NM || inputrec->eI == IntegrationAlgorithm::TPI)
1105 globalState = std::make_unique<t_state>();
1107 broadcastStateWithoutDynamics(
1108 cr->mpiDefaultCommunicator, DOMAINDECOMP(cr), PAR(cr), globalState.get());
1111 /* A parallel command line option consistency check that we can
1112 only do after any threads have started. */
1114 && (domdecOptions.numCells[XX] > 1 || domdecOptions.numCells[YY] > 1
1115 || domdecOptions.numCells[ZZ] > 1 || domdecOptions.numPmeRanks > 0))
1118 "The -dd or -npme option request a parallel simulation, "
1120 "but %s was compiled without threads or MPI enabled",
1121 output_env_get_program_display_name(oenv));
1122 #elif GMX_THREAD_MPI
1123 "but the number of MPI-threads (option -ntmpi) is not set or is 1");
1125 "but %s was not started through mpirun/mpiexec or only one rank was requested "
1126 "through mpirun/mpiexec",
1127 output_env_get_program_display_name(oenv));
1131 if (doRerun && (EI_ENERGY_MINIMIZATION(inputrec->eI) || IntegrationAlgorithm::NM == inputrec->eI))
1134 "The .mdp file specified an energy mininization or normal mode algorithm, and "
1135 "these are not compatible with mdrun -rerun");
1138 /* NMR restraints must be initialized before load_checkpoint,
1139 * since with time averaging the history is added to t_state.
1140 * For proper consistency check we therefore need to extend
1142 * So the PME-only nodes (if present) will also initialize
1143 * the distance restraints.
1146 /* This needs to be called before read_checkpoint to extend the state */
1147 t_disresdata* disresdata;
1148 snew(disresdata, 1);
1152 DisResRunMode::MDRun,
1153 MASTER(cr) ? DDRole::Master : DDRole::Agent,
1154 PAR(cr) ? NumRanks::Multiple : NumRanks::Single,
1159 replExParams.exchangeInterval > 0);
1161 if (gmx_mtop_ftype_count(mtop, F_ORIRES) > 0 && isSimulationMasterRank)
1163 extendStateWithOriresHistory(mtop, *inputrec, globalState.get());
1166 auto deform = prepareBoxDeformation(globalState != nullptr ? globalState->box : box,
1167 MASTER(cr) ? DDRole::Master : DDRole::Agent,
1168 PAR(cr) ? NumRanks::Multiple : NumRanks::Single,
1169 cr->mpi_comm_mygroup,
1173 /* We have to remember the generation's first step before reading checkpoint.
1174 This way, we can report to the F@H core both the generation's first step
1175 and the restored first step, thus making it able to distinguish between
1176 an interruption/resume and start of the n-th generation simulation.
1177 Having this information, the F@H core can correctly calculate and report
1180 int gen_first_step = 0;
1183 gen_first_step = inputrec->init_step;
1187 ObservablesHistory observablesHistory = {};
1189 auto modularSimulatorCheckpointData = std::make_unique<ReadCheckpointDataHolder>();
1190 if (startingBehavior != StartingBehavior::NewSimulation)
1192 /* Check if checkpoint file exists before doing continuation.
1193 * This way we can use identical input options for the first and subsequent runs...
1195 if (mdrunOptions.numStepsCommandline > -2)
1197 /* Temporarily set the number of steps to unlimited to avoid
1198 * triggering the nsteps check in load_checkpoint().
1199 * This hack will go away soon when the -nsteps option is removed.
1201 inputrec->nsteps = -1;
1204 // Finish applying initial simulation state information from external sources on all ranks.
1205 // Reconcile checkpoint file data with Mdrunner state established up to this point.
1206 applyLocalState(*inputHolder_.get(),
1209 domdecOptions.numCells,
1212 &observablesHistory,
1213 mdrunOptions.reproducible,
1214 mdModules_->notifiers(),
1215 modularSimulatorCheckpointData.get(),
1216 useModularSimulator);
1217 // TODO: (#3652) Synchronize filesystem state, SimulationInput contents, and program
1219 // on all code paths.
1220 // Write checkpoint or provide hook to update SimulationInput.
1221 // If there was a checkpoint file, SimulationInput contains more information
1222 // than if there wasn't. At this point, we have synchronized the in-memory
1223 // state with the filesystem state only for restarted simulations. We should
1224 // be calling applyLocalState unconditionally and expect that the completeness
1225 // of SimulationInput is not dependent on its creation method.
1227 if (startingBehavior == StartingBehavior::RestartWithAppending && logFileHandle)
1229 // Now we can start normal logging to the truncated log file.
1230 fplog = gmx_fio_getfp(logFileHandle);
1231 prepareLogAppending(fplog);
1232 logOwner = buildLogger(fplog, MASTER(cr));
1233 mdlog = logOwner.logger();
1240 fcRegisterSteps(inputrec->nsteps + inputrec->init_step, gen_first_step);
1244 if (mdrunOptions.numStepsCommandline > -2)
1249 "The -nsteps functionality is deprecated, and may be removed in a future "
1251 "Consider using gmx convert-tpr -nsteps or changing the appropriate .mdp "
1254 /* override nsteps with value set on the commandline */
1255 override_nsteps_cmdline(mdlog, mdrunOptions.numStepsCommandline, inputrec.get());
1257 if (isSimulationMasterRank)
1259 copy_mat(globalState->box, box);
1264 gmx_bcast(sizeof(box), box, cr->mpiDefaultCommunicator);
1267 if (inputrec->cutoff_scheme != CutoffScheme::Verlet)
1270 "This group-scheme .tpr file can no longer be run by mdrun. Please update to the "
1271 "Verlet scheme, or use an earlier version of GROMACS if necessary.");
1273 /* Update rlist and nstlist. */
1274 /* Note: prepare_verlet_scheme is calling increaseNstlist(...), which (while attempting to
1275 * increase rlist) tries to check if the newly chosen value fits with the DD scheme. As this is
1276 * run before any DD scheme is set up, this check is never executed. See #3334 for more details.
1278 prepare_verlet_scheme(fplog,
1284 useGpuForNonbonded || (emulateGpuNonbonded == EmulateGpuNonbonded::Yes),
1287 // We need to decide on update groups early, as this affects
1288 // inter-domain communication distances.
1289 auto updateGroupingsPerMoleculeType = makeUpdateGroupingsPerMoleculeType(mtop);
1290 const real maxUpdateGroupRadius = computeMaxUpdateGroupRadius(
1291 mtop, updateGroupingsPerMoleculeType, maxReferenceTemperature(*inputrec));
1292 const real cutoffMargin = std::sqrt(max_cutoff2(inputrec->pbcType, box)) - inputrec->rlist;
1293 UpdateGroups updateGroups = makeUpdateGroups(mdlog,
1294 std::move(updateGroupingsPerMoleculeType),
1295 maxUpdateGroupRadius,
1296 useDomainDecomposition,
1297 systemHasConstraintsOrVsites(mtop),
1302 const bool haveFrozenAtoms = inputrecFrozenAtoms(inputrec.get());
1304 useGpuForUpdate = decideWhetherToUseGpuForUpdate(useDomainDecomposition,
1305 updateGroups.useUpdateGroups(),
1307 domdecOptions.numPmeRanks > 0,
1313 doEssentialDynamics,
1314 gmx_mtop_ftype_count(mtop, F_ORIRES) > 0,
1320 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1322 bool useGpuDirectHalo = false;
1324 if (useGpuForNonbonded)
1326 // cr->npmenodes is not yet initialized.
1327 // domdecOptions.numPmeRanks == -1 results in 0 separate PME ranks when useGpuForNonbonded is true.
1328 // Todo: remove this assumption later once auto mode has support for separate PME rank
1329 const int numPmeRanks = domdecOptions.numPmeRanks > 0 ? domdecOptions.numPmeRanks : 0;
1330 bool havePPDomainDecomposition = (cr->sizeOfDefaultCommunicator - numPmeRanks) > 1;
1331 useGpuDirectHalo = decideWhetherToUseGpuForHalo(devFlags,
1332 havePPDomainDecomposition,
1334 useModularSimulator,
1336 EI_ENERGY_MINIMIZATION(inputrec->eI));
1339 // This builder is necessary while we have multi-part construction
1340 // of DD. Before DD is constructed, we use the existence of
1341 // the builder object to indicate that further construction of DD
1343 std::unique_ptr<DomainDecompositionBuilder> ddBuilder;
1344 if (useDomainDecomposition)
1346 // P2P GPU comm + GPU update leads to case in which we enqueue async work for multiple
1347 // timesteps. DLB needs to be disabled in that case
1348 const bool directGpuCommUsedWithGpuUpdate = GMX_THREAD_MPI && useGpuDirectHalo && useGpuForUpdate;
1349 ddBuilder = std::make_unique<DomainDecompositionBuilder>(
1356 mdModules_->notifiers(),
1358 updateGroups.updateGroupingPerMoleculeType(),
1359 updateGroups.useUpdateGroups(),
1360 updateGroups.maxUpdateGroupRadius(),
1361 positionsFromStatePointer(globalState.get()),
1364 directGpuCommUsedWithGpuUpdate);
1368 /* PME, if used, is done on all nodes with 1D decomposition */
1369 cr->nnodes = cr->sizeOfDefaultCommunicator;
1370 cr->sim_nodeid = cr->rankInDefaultCommunicator;
1371 cr->nodeid = cr->rankInDefaultCommunicator;
1373 cr->duty = (DUTY_PP | DUTY_PME);
1375 if (inputrec->pbcType == PbcType::Screw)
1377 gmx_fatal(FARGS, "pbc=screw is only implemented with domain decomposition");
1381 // Produce the task assignment for this rank - done after DD is constructed
1382 GpuTaskAssignments gpuTaskAssignments = GpuTaskAssignmentsBuilder::build(
1384 userGpuTaskAssignment,
1386 simulationCommunicator,
1394 thisRankHasDuty(cr, DUTY_PP),
1395 // TODO cr->duty & DUTY_PME should imply that a PME
1396 // algorithm is active, but currently does not.
1397 EEL_PME(inputrec->coulombtype) && thisRankHasDuty(cr, DUTY_PME));
1399 // Get the device handles for the modules, nullptr when no task is assigned.
1401 DeviceInformation* deviceInfo = gpuTaskAssignments.initDevice(&deviceId);
1403 // timing enabling - TODO put this in gpu_utils (even though generally this is just option handling?)
1404 bool useTiming = true;
1408 /* WARNING: CUDA timings are incorrect with multiple streams.
1409 * This is the main reason why they are disabled by default.
1411 // TODO: Consider turning on by default when we can detect nr of streams.
1412 useTiming = (getenv("GMX_ENABLE_GPU_TIMING") != nullptr);
1414 else if (GMX_GPU_OPENCL)
1416 useTiming = (getenv("GMX_DISABLE_GPU_TIMING") == nullptr);
1419 // TODO Currently this is always built, yet DD partition code
1420 // checks if it is built before using it. Probably it should
1421 // become an MDModule that is made only when another module
1422 // requires it (e.g. pull, CompEl, density fitting), so that we
1423 // don't update the local atom sets unilaterally every step.
1424 LocalAtomSetManager atomSets;
1427 // TODO Pass the GPU streams to ddBuilder to use in buffer
1428 // transfers (e.g. halo exchange)
1429 cr->dd = ddBuilder->build(&atomSets);
1430 // The builder's job is done, so destruct it
1431 ddBuilder.reset(nullptr);
1432 // Note that local state still does not exist yet.
1434 // Ensure that all atoms within the same update group are in the
1435 // same periodic image. Otherwise, a simulation that did not use
1436 // update groups (e.g. a single-rank simulation) cannot always be
1437 // correctly restarted in a way that does use update groups
1438 // (e.g. a multi-rank simulation).
1439 if (isSimulationMasterRank)
1441 const bool useUpdateGroups = cr->dd ? ddUsesUpdateGroups(*cr->dd) : false;
1442 if (useUpdateGroups)
1444 putUpdateGroupAtomsInSamePeriodicImage(*cr->dd, mtop, globalState->box, globalState->x);
1448 const bool printHostName = (cr->nnodes > 1);
1449 gpuTaskAssignments.reportGpuUsage(mdlog, printHostName, useGpuForBonded, pmeRunMode, useGpuForUpdate);
1451 const bool disableNonbondedCalculation = (getenv("GMX_NO_NONBONDED") != nullptr);
1452 if (disableNonbondedCalculation)
1454 /* turn off non-bonded calculations */
1455 GMX_LOG(mdlog.warning)
1458 "Found environment variable GMX_NO_NONBONDED.\n"
1459 "Disabling nonbonded calculations.");
1462 MdrunScheduleWorkload runScheduleWork;
1464 // Also populates the simulation constant workload description.
1465 // Note: currently the default duty is DUTY_PP | DUTY_PME for all simulations, including those without PME,
1466 // so this boolean is sufficient on all ranks to determine whether separate PME ranks are used,
1467 // but this will no longer be the case if cr->duty is changed for !EEL_PME(fr->ic->eeltype).
1468 const bool haveSeparatePmeRank = (!thisRankHasDuty(cr, DUTY_PP) || !thisRankHasDuty(cr, DUTY_PME));
1469 runScheduleWork.simulationWork = createSimulationWorkload(*inputrec,
1470 disableNonbondedCalculation,
1472 havePPDomainDecomposition(cr),
1473 haveSeparatePmeRank,
1480 std::unique_ptr<DeviceStreamManager> deviceStreamManager = nullptr;
1482 if (deviceInfo != nullptr)
1484 if (DOMAINDECOMP(cr) && thisRankHasDuty(cr, DUTY_PP))
1486 dd_setup_dlb_resource_sharing(cr, deviceId);
1488 deviceStreamManager = std::make_unique<DeviceStreamManager>(
1489 *deviceInfo, havePPDomainDecomposition(cr), runScheduleWork.simulationWork, useTiming);
1492 // If the user chose a task assignment, give them some hints
1493 // where appropriate.
1494 if (!userGpuTaskAssignment.empty())
1496 gpuTaskAssignments.logPerformanceHints(mdlog, numAvailableDevices);
1501 /* After possible communicator splitting in make_dd_communicators.
1502 * we can set up the intra/inter node communication.
1504 gmx_setup_nodecomm(fplog, cr);
1510 GMX_LOG(mdlog.warning)
1512 .appendTextFormatted(
1513 "This is simulation %d out of %d running as a composite GROMACS\n"
1514 "multi-simulation job. Setup for this simulation:\n",
1515 ms->simulationIndex_,
1516 ms->numSimulations_);
1518 GMX_LOG(mdlog.warning)
1519 .appendTextFormatted("Using %d MPI %s\n",
1522 cr->nnodes == 1 ? "thread" : "threads"
1524 cr->nnodes == 1 ? "process" : "processes"
1530 // If mdrun -pin auto honors any affinity setting that already
1531 // exists. If so, it is nice to provide feedback about whether
1532 // that existing affinity setting was from OpenMP or something
1533 // else, so we run this code both before and after we initialize
1534 // the OpenMP support.
1535 gmx_check_thread_affinity_set(mdlog, &hw_opt, hwinfo_->nthreads_hw_avail, FALSE);
1536 /* Check and update the number of OpenMP threads requested */
1537 checkAndUpdateRequestedNumOpenmpThreads(
1538 &hw_opt, *hwinfo_, cr, ms, physicalNodeComm.size_, pmeRunMode, mtop, *inputrec);
1540 gmx_omp_nthreads_init(mdlog,
1542 hwinfo_->nthreads_hw_avail,
1543 physicalNodeComm.size_,
1544 hw_opt.nthreads_omp,
1545 hw_opt.nthreads_omp_pme,
1546 !thisRankHasDuty(cr, DUTY_PP));
1548 const bool bEnableFPE = gmxShouldEnableFPExceptions();
1549 // FIXME - reconcile with gmx_feenableexcept() call from CommandLineModuleManager::run()
1552 gmx_feenableexcept();
1555 /* Now that we know the setup is consistent, check for efficiency */
1556 check_resource_division_efficiency(
1557 hwinfo_, gpuTaskAssignments.thisRankHasAnyGpuTask(), mdrunOptions.ntompOptionIsSet, cr, mdlog);
1559 /* getting number of PP/PME threads on this MPI / tMPI rank.
1560 PME: env variable should be read only on one node to make sure it is
1561 identical everywhere;
1563 const int numThreadsOnThisRank = thisRankHasDuty(cr, DUTY_PP)
1564 ? gmx_omp_nthreads_get(ModuleMultiThread::Nonbonded)
1565 : gmx_omp_nthreads_get(ModuleMultiThread::Pme);
1566 checkHardwareOversubscription(
1567 numThreadsOnThisRank, cr->nodeid, *hwinfo_->hardwareTopology, physicalNodeComm, mdlog);
1569 // Enable Peer access between GPUs where available
1570 // Only for DD, only master PP rank needs to perform setup, and only if thread MPI plus
1571 // any of the GPU communication features are active.
1572 if (DOMAINDECOMP(cr) && MASTER(cr) && thisRankHasDuty(cr, DUTY_PP) && GMX_THREAD_MPI
1573 && (runScheduleWork.simulationWork.useGpuHaloExchange
1574 || runScheduleWork.simulationWork.useGpuPmePpCommunication))
1576 setupGpuDevicePeerAccess(gpuTaskAssignments.deviceIdsAssigned(), mdlog);
1579 if (hw_opt.threadAffinity != ThreadAffinity::Off)
1581 /* Before setting affinity, check whether the affinity has changed
1582 * - which indicates that probably the OpenMP library has changed it
1583 * since we first checked).
1585 gmx_check_thread_affinity_set(mdlog, &hw_opt, hwinfo_->nthreads_hw_avail, TRUE);
1587 int numThreadsOnThisNode, intraNodeThreadOffset;
1588 analyzeThreadsOnThisNode(
1589 physicalNodeComm, numThreadsOnThisRank, &numThreadsOnThisNode, &intraNodeThreadOffset);
1591 /* Set the CPU affinity */
1592 gmx_set_thread_affinity(mdlog,
1595 *hwinfo_->hardwareTopology,
1596 numThreadsOnThisRank,
1597 numThreadsOnThisNode,
1598 intraNodeThreadOffset,
1602 if (mdrunOptions.timingOptions.resetStep > -1)
1607 "The -resetstep functionality is deprecated, and may be removed in a "
1610 std::unique_ptr<gmx_wallcycle> wcycle =
1611 wallcycle_init(fplog, mdrunOptions.timingOptions.resetStep, cr);
1615 /* Master synchronizes its value of reset_counters with all nodes
1616 * including PME only nodes */
1617 int64_t reset_counters = wcycle_get_reset_counters(wcycle.get());
1618 gmx_bcast(sizeof(reset_counters), &reset_counters, cr->mpi_comm_mysim);
1619 wcycle_set_reset_counters(wcycle.get(), reset_counters);
1622 // Membrane embedding must be initialized before we call init_forcerec()
1623 membedHolder.initializeMembed(fplog,
1630 &mdrunOptions.checkpointOptions.period);
1632 const bool thisRankHasPmeGpuTask = gpuTaskAssignments.thisRankHasPmeGpuTask();
1633 std::unique_ptr<MDAtoms> mdAtoms;
1634 std::unique_ptr<VirtualSitesHandler> vsite;
1637 if (thisRankHasDuty(cr, DUTY_PP))
1639 setupNotifier.notify(*cr);
1640 setupNotifier.notify(&atomSets);
1641 setupNotifier.notify(mtop);
1642 setupNotifier.notify(inputrec->pbcType);
1643 setupNotifier.notify(SimulationTimeStep{ inputrec->delta_t });
1644 /* Initiate forcerecord */
1645 fr = std::make_unique<t_forcerec>();
1646 fr->forceProviders = mdModules_->initForceProviders();
1647 init_forcerec(fplog,
1649 runScheduleWork.simulationWork,
1655 opt2fn("-table", filenames.size(), filenames.data()),
1656 opt2fn("-tablep", filenames.size(), filenames.data()),
1657 opt2fns("-tableb", filenames.size(), filenames.data()),
1659 // Dirty hack, for fixing disres and orires should be made mdmodules
1660 fr->fcdata->disres = disresdata;
1661 if (gmx_mtop_ftype_count(mtop, F_ORIRES) > 0)
1663 fr->fcdata->orires = std::make_unique<t_oriresdata>(
1664 fplog, mtop, *inputrec, ms, globalState.get(), &atomSets);
1667 // Save a handle to device stream manager to use elsewhere in the code
1668 // TODO: Forcerec is not a correct place to store it.
1669 fr->deviceStreamManager = deviceStreamManager.get();
1671 if (runScheduleWork.simulationWork.useGpuPmePpCommunication && !thisRankHasDuty(cr, DUTY_PME))
1674 deviceStreamManager != nullptr,
1675 "GPU device stream manager should be valid in order to use PME-PP direct "
1678 deviceStreamManager->streamIsValid(DeviceStreamType::PmePpTransfer),
1679 "GPU PP-PME stream should be valid in order to use GPU PME-PP direct "
1681 fr->pmePpCommGpu = std::make_unique<gmx::PmePpCommGpu>(
1684 &cr->dd->pmeForceReceiveBuffer,
1685 deviceStreamManager->context(),
1686 deviceStreamManager->stream(DeviceStreamType::PmePpTransfer));
1689 fr->nbv = Nbnxm::init_nb_verlet(mdlog,
1694 runScheduleWork.simulationWork.useGpuNonbonded,
1695 deviceStreamManager.get(),
1699 // TODO: Move the logic below to a GPU bonded builder
1700 if (runScheduleWork.simulationWork.useGpuBonded)
1702 GMX_RELEASE_ASSERT(deviceStreamManager != nullptr,
1703 "GPU device stream manager should be valid in order to use GPU "
1704 "version of bonded forces.");
1705 fr->listedForcesGpu = std::make_unique<ListedForcesGpu>(
1707 fr->ic->epsfac * fr->fudgeQQ,
1708 deviceStreamManager->context(),
1709 deviceStreamManager->bondedStream(havePPDomainDecomposition(cr)),
1713 /* Initialize the mdAtoms structure.
1714 * mdAtoms is not filled with atom data,
1715 * as this can not be done now with domain decomposition.
1717 mdAtoms = makeMDAtoms(fplog, mtop, *inputrec, thisRankHasPmeGpuTask);
1718 if (globalState && thisRankHasPmeGpuTask)
1720 // The pinning of coordinates in the global state object works, because we only use
1721 // PME on GPU without DD or on a separate PME rank, and because the local state pointer
1722 // points to the global state object without DD.
1723 // FIXME: MD and EM separately set up the local state - this should happen in the same
1724 // function, which should also perform the pinning.
1725 changePinningPolicy(&globalState->x, pme_get_pinning_policy());
1728 /* Initialize the virtual site communication */
1729 vsite = makeVirtualSitesHandler(
1730 mtop, cr, fr->pbcType, updateGroups.updateGroupingPerMoleculeType());
1732 calc_shifts(box, fr->shift_vec);
1734 /* With periodic molecules the charge groups should be whole at start up
1735 * and the virtual sites should not be far from their proper positions.
1737 if (!inputrec->bContinuation && MASTER(cr)
1738 && !(inputrec->pbcType != PbcType::No && inputrec->bPeriodicMols))
1740 /* Make molecules whole at start of run */
1741 if (fr->pbcType != PbcType::No)
1743 do_pbc_first_mtop(fplog, inputrec->pbcType, box, &mtop, globalState->x.rvec_array());
1747 /* Correct initial vsite positions are required
1748 * for the initial distribution in the domain decomposition
1749 * and for the initial shell prediction.
1751 constructVirtualSitesGlobal(mtop, globalState->x);
1754 // Make the DD reverse topology, now that any vsites that are present are available
1755 if (DOMAINDECOMP(cr))
1757 dd_make_reverse_top(fplog, cr->dd, mtop, vsite.get(), *inputrec, domdecOptions.ddBondedChecking);
1760 if (EEL_PME(fr->ic->eeltype) || EVDW_PME(fr->ic->vdwtype))
1762 ewaldcoeff_q = fr->ic->ewaldcoeff_q;
1763 ewaldcoeff_lj = fr->ic->ewaldcoeff_lj;
1768 /* This is a PME only node */
1770 GMX_ASSERT(globalState == nullptr,
1771 "We don't need the state on a PME only rank and expect it to be unitialized");
1773 ewaldcoeff_q = calc_ewaldcoeff_q(inputrec->rcoulomb, inputrec->ewald_rtol);
1774 ewaldcoeff_lj = calc_ewaldcoeff_lj(inputrec->rvdw, inputrec->ewald_rtol_lj);
1777 gmx_pme_t* sepPmeData = nullptr;
1778 // This reference hides the fact that PME data is owned by runner on PME-only ranks and by forcerec on other ranks
1779 GMX_ASSERT(thisRankHasDuty(cr, DUTY_PP) == (fr != nullptr),
1780 "Double-checking that only PME-only ranks have no forcerec");
1781 gmx_pme_t*& pmedata = fr ? fr->pmedata : sepPmeData;
1783 // TODO should live in ewald module once its testing is improved
1785 // Later, this program could contain kernels that might be later
1786 // re-used as auto-tuning progresses, or subsequent simulations
1788 PmeGpuProgramStorage pmeGpuProgram;
1789 if (thisRankHasPmeGpuTask)
1792 (deviceStreamManager != nullptr),
1793 "GPU device stream manager should be initialized in order to use GPU for PME.");
1794 GMX_RELEASE_ASSERT((deviceInfo != nullptr),
1795 "GPU device should be initialized in order to use GPU for PME.");
1796 pmeGpuProgram = buildPmeGpuProgram(deviceStreamManager->context());
1799 /* Initiate PME if necessary,
1800 * either on all nodes or on dedicated PME nodes only. */
1801 if (EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype))
1803 if (mdAtoms && mdAtoms->mdatoms())
1805 nChargePerturbed = mdAtoms->mdatoms()->nChargePerturbed;
1806 if (EVDW_PME(inputrec->vdwtype))
1808 nTypePerturbed = mdAtoms->mdatoms()->nTypePerturbed;
1811 if (cr->npmenodes > 0)
1813 /* The PME only nodes need to know nChargePerturbed(FEP on Q) and nTypePerturbed(FEP on LJ)*/
1814 gmx_bcast(sizeof(nChargePerturbed), &nChargePerturbed, cr->mpi_comm_mysim);
1815 gmx_bcast(sizeof(nTypePerturbed), &nTypePerturbed, cr->mpi_comm_mysim);
1818 if (thisRankHasDuty(cr, DUTY_PME))
1822 // TODO: This should be in the builder.
1823 GMX_RELEASE_ASSERT(!runScheduleWork.simulationWork.useGpuPme
1824 || (deviceStreamManager != nullptr),
1825 "Device stream manager should be valid in order to use GPU "
1828 !runScheduleWork.simulationWork.useGpuPme
1829 || deviceStreamManager->streamIsValid(DeviceStreamType::Pme),
1830 "GPU PME stream should be valid in order to use GPU version of PME.");
1832 const DeviceContext* deviceContext = runScheduleWork.simulationWork.useGpuPme
1833 ? &deviceStreamManager->context()
1835 const DeviceStream* pmeStream =
1836 runScheduleWork.simulationWork.useGpuPme
1837 ? &deviceStreamManager->stream(DeviceStreamType::Pme)
1840 pmedata = gmx_pme_init(cr,
1841 getNumPmeDomains(cr->dd),
1843 nChargePerturbed != 0,
1844 nTypePerturbed != 0,
1845 mdrunOptions.reproducible,
1848 gmx_omp_nthreads_get(ModuleMultiThread::Pme),
1853 pmeGpuProgram.get(),
1856 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1861 if (EI_DYNAMICS(inputrec->eI))
1863 /* Turn on signal handling on all nodes */
1865 * (A user signal from the PME nodes (if any)
1866 * is communicated to the PP nodes.
1868 signal_handler_install();
1871 pull_t* pull_work = nullptr;
1872 if (thisRankHasDuty(cr, DUTY_PP))
1874 /* Assumes uniform use of the number of OpenMP threads */
1875 walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(ModuleMultiThread::Default));
1877 if (inputrec->bPull)
1879 /* Initialize pull code */
1880 pull_work = init_pull(fplog,
1881 inputrec->pull.get(),
1886 inputrec->fepvals->init_lambda);
1887 if (inputrec->pull->bXOutAverage || inputrec->pull->bFOutAverage)
1889 initPullHistory(pull_work, &observablesHistory);
1891 if (EI_DYNAMICS(inputrec->eI) && MASTER(cr))
1893 init_pull_output_files(pull_work, filenames.size(), filenames.data(), oenv, startingBehavior);
1897 std::unique_ptr<EnforcedRotation> enforcedRotation;
1900 /* Initialize enforced rotation code */
1901 enforcedRotation = init_rot(fplog,
1914 t_swap* swap = nullptr;
1915 if (inputrec->eSwapCoords != SwapType::No)
1917 /* Initialize ion swapping code */
1918 swap = init_swapcoords(fplog,
1920 opt2fn_master("-swap", filenames.size(), filenames.data(), cr),
1923 &observablesHistory,
1931 /* Let makeConstraints know whether we have essential dynamics constraints. */
1932 auto constr = makeConstraints(mtop,
1935 doEssentialDynamics,
1938 updateGroups.useUpdateGroups(),
1944 /* Energy terms and groups */
1945 gmx_enerdata_t enerd(mtop.groups.groups[SimulationAtomGroupType::EnergyOutput].size(),
1946 inputrec->fepvals->n_lambda);
1948 // cos acceleration is only supported by md, but older tpr
1949 // files might still combine it with other integrators
1950 GMX_RELEASE_ASSERT(inputrec->cos_accel == 0.0 || inputrec->eI == IntegrationAlgorithm::MD,
1951 "cos_acceleration is only supported by integrator=md");
1953 /* Kinetic energy data */
1954 gmx_ekindata_t ekind(inputrec->opts.ngtc,
1955 inputrec->cos_accel,
1956 gmx_omp_nthreads_get(ModuleMultiThread::Update));
1958 /* Set up interactive MD (IMD) */
1959 auto imdSession = makeImdSession(inputrec.get(),
1966 MASTER(cr) ? globalState->x : gmx::ArrayRef<gmx::RVec>(),
1970 mdrunOptions.imdOptions,
1973 if (DOMAINDECOMP(cr))
1975 GMX_RELEASE_ASSERT(fr, "fr was NULL while cr->duty was DUTY_PP");
1976 /* This call is not included in init_domain_decomposition
1977 * because fr->atomInfoForEachMoleculeBlock is set later.
1979 makeBondedLinks(cr->dd, mtop, fr->atomInfoForEachMoleculeBlock);
1982 if (runScheduleWork.simulationWork.useGpuBufferOps)
1984 fr->gpuForceReduction[gmx::AtomLocality::Local] = std::make_unique<gmx::GpuForceReduction>(
1985 deviceStreamManager->context(),
1986 deviceStreamManager->stream(gmx::DeviceStreamType::NonBondedLocal),
1988 fr->gpuForceReduction[gmx::AtomLocality::NonLocal] = std::make_unique<gmx::GpuForceReduction>(
1989 deviceStreamManager->context(),
1990 deviceStreamManager->stream(gmx::DeviceStreamType::NonBondedNonLocal),
1994 std::unique_ptr<gmx::StatePropagatorDataGpu> stateGpu;
1995 if (gpusWereDetected
1996 && ((runScheduleWork.simulationWork.useGpuPme && thisRankHasDuty(cr, DUTY_PME))
1997 || runScheduleWork.simulationWork.useGpuBufferOps))
1999 GpuApiCallBehavior transferKind =
2000 (inputrec->eI == IntegrationAlgorithm::MD && !doRerun && !useModularSimulator)
2001 ? GpuApiCallBehavior::Async
2002 : GpuApiCallBehavior::Sync;
2003 GMX_RELEASE_ASSERT(deviceStreamManager != nullptr,
2004 "GPU device stream manager should be initialized to use GPU.");
2005 stateGpu = std::make_unique<gmx::StatePropagatorDataGpu>(
2006 *deviceStreamManager, transferKind, pme_gpu_get_block_size(fr->pmedata), wcycle.get());
2007 fr->stateGpu = stateGpu.get();
2010 GMX_ASSERT(stopHandlerBuilder_, "Runner must provide StopHandlerBuilder to simulator.");
2011 SimulatorBuilder simulatorBuilder;
2013 simulatorBuilder.add(SimulatorStateData(globalState.get(), &observablesHistory, &enerd, &ekind));
2014 simulatorBuilder.add(std::move(membedHolder));
2015 simulatorBuilder.add(std::move(stopHandlerBuilder_));
2016 simulatorBuilder.add(SimulatorConfig(mdrunOptions, startingBehavior, &runScheduleWork));
2019 simulatorBuilder.add(SimulatorEnv(fplog, cr, ms, mdlog, oenv));
2020 simulatorBuilder.add(Profiling(&nrnb, walltime_accounting, wcycle.get()));
2021 simulatorBuilder.add(ConstraintsParam(
2022 constr.get(), enforcedRotation ? enforcedRotation->getLegacyEnfrot() : nullptr, vsite.get()));
2023 // TODO: Separate `fr` to a separate add, and make the `build` handle the coupling sensibly.
2024 simulatorBuilder.add(LegacyInput(
2025 static_cast<int>(filenames.size()), filenames.data(), inputrec.get(), fr.get()));
2026 simulatorBuilder.add(ReplicaExchangeParameters(replExParams));
2027 simulatorBuilder.add(InteractiveMD(imdSession.get()));
2028 simulatorBuilder.add(SimulatorModules(mdModules_->outputProvider(), mdModules_->notifiers()));
2029 simulatorBuilder.add(CenterOfMassPulling(pull_work));
2030 // Todo move to an MDModule
2031 simulatorBuilder.add(IonSwapping(swap));
2032 simulatorBuilder.add(TopologyData(mtop, mdAtoms.get()));
2033 simulatorBuilder.add(BoxDeformationHandle(deform.get()));
2034 simulatorBuilder.add(std::move(modularSimulatorCheckpointData));
2036 // build and run simulator object based on user-input
2037 auto simulator = simulatorBuilder.build(useModularSimulator);
2040 if (fr->pmePpCommGpu)
2042 // destroy object since it is no longer required. (This needs to be done while the GPU context still exists.)
2043 fr->pmePpCommGpu.reset();
2046 if (inputrec->bPull)
2048 finish_pull(pull_work);
2050 finish_swapcoords(swap);
2054 GMX_RELEASE_ASSERT(pmedata, "pmedata was NULL while cr->duty was not DUTY_PP");
2056 walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(ModuleMultiThread::Pme));
2057 gmx_pmeonly(pmedata,
2061 walltime_accounting,
2064 runScheduleWork.simulationWork.useGpuPmePpCommunication,
2065 deviceStreamManager.get());
2068 wallcycle_stop(wcycle.get(), WallCycleCounter::Run);
2070 /* Finish up, write some stuff
2071 * if rerunMD, don't write last frame again
2079 walltime_accounting,
2080 fr ? fr->nbv.get() : nullptr,
2082 EI_DYNAMICS(inputrec->eI) && !isMultiSim(ms));
2085 deviceStreamManager.reset(nullptr);
2089 gmx_pme_destroy(pmedata);
2093 // FIXME: this is only here to manually unpin mdAtoms->chargeA_ and state->x,
2094 // before we destroy the GPU context(s)
2095 // Pinned buffers are associated with contexts in CUDA.
2096 // As soon as we destroy GPU contexts after mdrunner() exits, these lines should go.
2097 mdAtoms.reset(nullptr);
2098 globalState.reset(nullptr);
2099 mdModules_.reset(nullptr); // destruct force providers here as they might also use the GPU
2100 fr.reset(nullptr); // destruct forcerec before gpu
2101 // TODO convert to C++ so we can get rid of these frees
2104 if (!hwinfo_->deviceInfoList.empty())
2106 /* stop the GPU profiler (only CUDA) */
2110 /* With tMPI we need to wait for all ranks to finish deallocation before
2111 * destroying the CUDA context as some tMPI ranks may be sharing
2114 * This is not a concern in OpenCL where we use one context per rank.
2116 * Note: it is safe to not call the barrier on the ranks which do not use GPU,
2117 * but it is easier and more futureproof to call it on the whole node.
2119 * Note that this function needs to be called even if GPUs are not used
2120 * in this run because the PME ranks have no knowledge of whether GPUs
2121 * are used or not, but all ranks need to enter the barrier below.
2122 * \todo Remove this physical node barrier after making sure
2123 * that it's not needed anymore (with a shared GPU run).
2127 physicalNodeComm.barrier();
2130 if (!devFlags.usingCudaAwareMpi)
2132 // Don't reset GPU in case of CUDA-AWARE MPI
2133 // UCX creates CUDA buffers which are cleaned-up as part of MPI_Finalize()
2134 // resetting the device before MPI_Finalize() results in crashes inside UCX
2135 releaseDevice(deviceInfo);
2138 /* Does what it says */
2139 print_date_and_time(fplog, cr->nodeid, "Finished mdrun", gmx_gettime());
2140 walltime_accounting_destroy(walltime_accounting);
2142 // Ensure log file content is written
2145 gmx_fio_flush(logFileHandle);
2148 /* Reset FPEs (important for unit tests) by disabling them. Assumes no
2149 * exceptions were enabled before function was called. */
2152 gmx_fedisableexcept();
2155 auto rc = static_cast<int>(gmx_get_stop_condition());
2158 /* we need to join all threads. The sub-threads join when they
2159 exit this function, but the master thread needs to be told to
2169 Mdrunner::~Mdrunner()
2171 // Clean up of the Manager.
2172 // This will end up getting called on every thread-MPI rank, which is unnecessary,
2173 // but okay as long as threads synchronize some time before adding or accessing
2174 // a new set of restraints.
2175 if (restraintManager_)
2177 restraintManager_->clear();
2178 GMX_ASSERT(restraintManager_->countRestraints() == 0,
2179 "restraints added during runner life time should be cleared at runner "
2184 void Mdrunner::addPotential(std::shared_ptr<gmx::IRestraintPotential> puller, const std::string& name)
2186 GMX_ASSERT(restraintManager_, "Mdrunner must have a restraint manager.");
2187 // Not sure if this should be logged through the md logger or something else,
2188 // but it is helpful to have some sort of INFO level message sent somewhere.
2189 // std::cout << "Registering restraint named " << name << std::endl;
2191 // When multiple restraints are used, it may be wasteful to register them separately.
2192 // Maybe instead register an entire Restraint Manager as a force provider.
2193 restraintManager_->addToSpec(std::move(puller), name);
2196 Mdrunner::Mdrunner(std::unique_ptr<MDModules> mdModules) : mdModules_(std::move(mdModules)) {}
2198 Mdrunner::Mdrunner(Mdrunner&&) noexcept = default;
2200 //NOLINTNEXTLINE(performance-noexcept-move-constructor) working around GCC bug 58265 in CentOS 7
2201 Mdrunner& Mdrunner::operator=(Mdrunner&& /*handle*/) noexcept(BUGFREE_NOEXCEPT_STRING) = default;
2203 class Mdrunner::BuilderImplementation
2206 BuilderImplementation() = delete;
2207 BuilderImplementation(std::unique_ptr<MDModules> mdModules, compat::not_null<SimulationContext*> context);
2208 ~BuilderImplementation();
2210 BuilderImplementation& setExtraMdrunOptions(const MdrunOptions& options,
2211 real forceWarningThreshold,
2212 StartingBehavior startingBehavior);
2214 void addHardwareDetectionResult(const gmx_hw_info_t* hwinfo);
2216 void addDomdec(const DomdecOptions& options);
2218 void addInput(SimulationInputHandle inputHolder);
2220 void addVerletList(int nstlist);
2222 void addReplicaExchange(const ReplicaExchangeParameters& params);
2224 void addNonBonded(const char* nbpu_opt);
2226 void addPME(const char* pme_opt_, const char* pme_fft_opt_);
2228 void addBondedTaskAssignment(const char* bonded_opt);
2230 void addUpdateTaskAssignment(const char* update_opt);
2232 void addHardwareOptions(const gmx_hw_opt_t& hardwareOptions);
2234 void addFilenames(ArrayRef<const t_filenm> filenames);
2236 void addOutputEnvironment(gmx_output_env_t* outputEnvironment);
2238 void addLogFile(t_fileio* logFileHandle);
2240 void addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder);
2245 // Default parameters copied from runner.h
2246 // \todo Clarify source(s) of default parameters.
2248 const char* nbpu_opt_ = nullptr;
2249 const char* pme_opt_ = nullptr;
2250 const char* pme_fft_opt_ = nullptr;
2251 const char* bonded_opt_ = nullptr;
2252 const char* update_opt_ = nullptr;
2254 MdrunOptions mdrunOptions_;
2256 DomdecOptions domdecOptions_;
2258 ReplicaExchangeParameters replicaExchangeParameters_;
2260 //! Command-line override for the duration of a neighbor list with the Verlet scheme.
2263 //! World communicator, used for hardware detection and task assignment
2264 MPI_Comm libraryWorldCommunicator_ = MPI_COMM_NULL;
2266 //! Multisim communicator handle.
2267 gmx_multisim_t* multiSimulation_;
2269 //! mdrun communicator
2270 MPI_Comm simulationCommunicator_ = MPI_COMM_NULL;
2272 //! Print a warning if any force is larger than this (in kJ/mol nm).
2273 real forceWarningThreshold_ = -1;
2275 //! Whether the simulation will start afresh, or restart with/without appending.
2276 StartingBehavior startingBehavior_ = StartingBehavior::NewSimulation;
2278 //! The modules that comprise the functionality of mdrun.
2279 std::unique_ptr<MDModules> mdModules_;
2281 //! Detected hardware.
2282 const gmx_hw_info_t* hwinfo_ = nullptr;
2284 //! \brief Parallelism information.
2285 gmx_hw_opt_t hardwareOptions_;
2287 //! filename options for simulation.
2288 ArrayRef<const t_filenm> filenames_;
2290 /*! \brief Handle to output environment.
2292 * \todo gmx_output_env_t needs lifetime management.
2294 gmx_output_env_t* outputEnvironment_ = nullptr;
2296 /*! \brief Non-owning handle to MD log file.
2298 * \todo Context should own output facilities for client.
2299 * \todo Improve log file handle management.
2301 * Code managing the FILE* relies on the ability to set it to
2302 * nullptr to check whether the filehandle is valid.
2304 t_fileio* logFileHandle_ = nullptr;
2307 * \brief Builder for simulation stop signal handler.
2309 std::unique_ptr<StopHandlerBuilder> stopHandlerBuilder_ = nullptr;
2312 * \brief Sources for initial simulation state.
2314 * See issue #3652 for near-term refinements to the SimulationInput interface.
2316 * See issue #3379 for broader discussion on API aspects of simulation inputs and outputs.
2318 SimulationInputHandle inputHolder_;
2321 Mdrunner::BuilderImplementation::BuilderImplementation(std::unique_ptr<MDModules> mdModules,
2322 compat::not_null<SimulationContext*> context) :
2323 mdModules_(std::move(mdModules))
2325 libraryWorldCommunicator_ = context->libraryWorldCommunicator_;
2326 simulationCommunicator_ = context->simulationCommunicator_;
2327 multiSimulation_ = context->multiSimulation_.get();
2330 Mdrunner::BuilderImplementation::~BuilderImplementation() = default;
2332 Mdrunner::BuilderImplementation&
2333 Mdrunner::BuilderImplementation::setExtraMdrunOptions(const MdrunOptions& options,
2334 const real forceWarningThreshold,
2335 const StartingBehavior startingBehavior)
2337 mdrunOptions_ = options;
2338 forceWarningThreshold_ = forceWarningThreshold;
2339 startingBehavior_ = startingBehavior;
2343 void Mdrunner::BuilderImplementation::addDomdec(const DomdecOptions& options)
2345 domdecOptions_ = options;
2348 void Mdrunner::BuilderImplementation::addVerletList(int nstlist)
2353 void Mdrunner::BuilderImplementation::addReplicaExchange(const ReplicaExchangeParameters& params)
2355 replicaExchangeParameters_ = params;
2358 Mdrunner Mdrunner::BuilderImplementation::build()
2360 auto newRunner = Mdrunner(std::move(mdModules_));
2362 newRunner.mdrunOptions = mdrunOptions_;
2363 newRunner.pforce = forceWarningThreshold_;
2364 newRunner.startingBehavior = startingBehavior_;
2365 newRunner.domdecOptions = domdecOptions_;
2367 // \todo determine an invariant to check or confirm that all gmx_hw_opt_t objects are valid
2368 newRunner.hw_opt = hardwareOptions_;
2370 // No invariant to check. This parameter exists to optionally override other behavior.
2371 newRunner.nstlist_cmdline = nstlist_;
2373 newRunner.replExParams = replicaExchangeParameters_;
2375 newRunner.filenames = filenames_;
2377 newRunner.libraryWorldCommunicator = libraryWorldCommunicator_;
2379 newRunner.simulationCommunicator = simulationCommunicator_;
2381 // nullptr is a valid value for the multisim handle
2382 newRunner.ms = multiSimulation_;
2386 newRunner.hwinfo_ = hwinfo_;
2390 GMX_THROW(gmx::APIError(
2391 "MdrunnerBuilder::addHardwareDetectionResult() is required before build()"));
2396 newRunner.inputHolder_ = std::move(inputHolder_);
2400 GMX_THROW(gmx::APIError("MdrunnerBuilder::addInput() is required before build()."));
2403 // \todo Clarify ownership and lifetime management for gmx_output_env_t
2404 // \todo Update sanity checking when output environment has clearly specified invariants.
2405 // Initialization and default values for oenv are not well specified in the current version.
2406 if (outputEnvironment_)
2408 newRunner.oenv = outputEnvironment_;
2412 GMX_THROW(gmx::APIError(
2413 "MdrunnerBuilder::addOutputEnvironment() is required before build()"));
2416 newRunner.logFileHandle = logFileHandle_;
2420 newRunner.nbpu_opt = nbpu_opt_;
2424 GMX_THROW(gmx::APIError("MdrunnerBuilder::addNonBonded() is required before build()"));
2427 if (pme_opt_ && pme_fft_opt_)
2429 newRunner.pme_opt = pme_opt_;
2430 newRunner.pme_fft_opt = pme_fft_opt_;
2434 GMX_THROW(gmx::APIError("MdrunnerBuilder::addElectrostatics() is required before build()"));
2439 newRunner.bonded_opt = bonded_opt_;
2443 GMX_THROW(gmx::APIError(
2444 "MdrunnerBuilder::addBondedTaskAssignment() is required before build()"));
2449 newRunner.update_opt = update_opt_;
2453 GMX_THROW(gmx::APIError(
2454 "MdrunnerBuilder::addUpdateTaskAssignment() is required before build() "));
2458 newRunner.restraintManager_ = std::make_unique<gmx::RestraintManager>();
2460 if (stopHandlerBuilder_)
2462 newRunner.stopHandlerBuilder_ = std::move(stopHandlerBuilder_);
2466 newRunner.stopHandlerBuilder_ = std::make_unique<StopHandlerBuilder>();
2472 void Mdrunner::BuilderImplementation::addHardwareDetectionResult(const gmx_hw_info_t* hwinfo)
2477 void Mdrunner::BuilderImplementation::addNonBonded(const char* nbpu_opt)
2479 nbpu_opt_ = nbpu_opt;
2482 void Mdrunner::BuilderImplementation::addPME(const char* pme_opt, const char* pme_fft_opt)
2485 pme_fft_opt_ = pme_fft_opt;
2488 void Mdrunner::BuilderImplementation::addBondedTaskAssignment(const char* bonded_opt)
2490 bonded_opt_ = bonded_opt;
2493 void Mdrunner::BuilderImplementation::addUpdateTaskAssignment(const char* update_opt)
2495 update_opt_ = update_opt;
2498 void Mdrunner::BuilderImplementation::addHardwareOptions(const gmx_hw_opt_t& hardwareOptions)
2500 hardwareOptions_ = hardwareOptions;
2503 void Mdrunner::BuilderImplementation::addFilenames(ArrayRef<const t_filenm> filenames)
2505 filenames_ = filenames;
2508 void Mdrunner::BuilderImplementation::addOutputEnvironment(gmx_output_env_t* outputEnvironment)
2510 outputEnvironment_ = outputEnvironment;
2513 void Mdrunner::BuilderImplementation::addLogFile(t_fileio* logFileHandle)
2515 logFileHandle_ = logFileHandle;
2518 void Mdrunner::BuilderImplementation::addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder)
2520 stopHandlerBuilder_ = std::move(builder);
2523 void Mdrunner::BuilderImplementation::addInput(SimulationInputHandle inputHolder)
2525 inputHolder_ = std::move(inputHolder);
2528 MdrunnerBuilder::MdrunnerBuilder(std::unique_ptr<MDModules> mdModules,
2529 compat::not_null<SimulationContext*> context) :
2530 impl_{ std::make_unique<Mdrunner::BuilderImplementation>(std::move(mdModules), context) }
2534 MdrunnerBuilder::~MdrunnerBuilder() = default;
2536 MdrunnerBuilder& MdrunnerBuilder::addHardwareDetectionResult(const gmx_hw_info_t* hwinfo)
2538 impl_->addHardwareDetectionResult(hwinfo);
2542 MdrunnerBuilder& MdrunnerBuilder::addSimulationMethod(const MdrunOptions& options,
2543 real forceWarningThreshold,
2544 const StartingBehavior startingBehavior)
2546 impl_->setExtraMdrunOptions(options, forceWarningThreshold, startingBehavior);
2550 MdrunnerBuilder& MdrunnerBuilder::addDomainDecomposition(const DomdecOptions& options)
2552 impl_->addDomdec(options);
2556 MdrunnerBuilder& MdrunnerBuilder::addNeighborList(int nstlist)
2558 impl_->addVerletList(nstlist);
2562 MdrunnerBuilder& MdrunnerBuilder::addReplicaExchange(const ReplicaExchangeParameters& params)
2564 impl_->addReplicaExchange(params);
2568 MdrunnerBuilder& MdrunnerBuilder::addNonBonded(const char* nbpu_opt)
2570 impl_->addNonBonded(nbpu_opt);
2574 MdrunnerBuilder& MdrunnerBuilder::addElectrostatics(const char* pme_opt, const char* pme_fft_opt)
2576 // The builder method may become more general in the future, but in this version,
2577 // parameters for PME electrostatics are both required and the only parameters
2579 if (pme_opt && pme_fft_opt)
2581 impl_->addPME(pme_opt, pme_fft_opt);
2586 gmx::InvalidInputError("addElectrostatics() arguments must be non-null pointers."));
2591 MdrunnerBuilder& MdrunnerBuilder::addBondedTaskAssignment(const char* bonded_opt)
2593 impl_->addBondedTaskAssignment(bonded_opt);
2597 MdrunnerBuilder& MdrunnerBuilder::addUpdateTaskAssignment(const char* update_opt)
2599 impl_->addUpdateTaskAssignment(update_opt);
2603 Mdrunner MdrunnerBuilder::build()
2605 return impl_->build();
2608 MdrunnerBuilder& MdrunnerBuilder::addHardwareOptions(const gmx_hw_opt_t& hardwareOptions)
2610 impl_->addHardwareOptions(hardwareOptions);
2614 MdrunnerBuilder& MdrunnerBuilder::addFilenames(ArrayRef<const t_filenm> filenames)
2616 impl_->addFilenames(filenames);
2620 MdrunnerBuilder& MdrunnerBuilder::addOutputEnvironment(gmx_output_env_t* outputEnvironment)
2622 impl_->addOutputEnvironment(outputEnvironment);
2626 MdrunnerBuilder& MdrunnerBuilder::addLogFile(t_fileio* logFileHandle)
2628 impl_->addLogFile(logFileHandle);
2632 MdrunnerBuilder& MdrunnerBuilder::addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder)
2634 impl_->addStopHandlerBuilder(std::move(builder));
2638 MdrunnerBuilder& MdrunnerBuilder::addInput(SimulationInputHandle input)
2640 impl_->addInput(std::move(input));
2644 MdrunnerBuilder::MdrunnerBuilder(MdrunnerBuilder&&) noexcept = default;
2646 MdrunnerBuilder& MdrunnerBuilder::operator=(MdrunnerBuilder&&) noexcept = default;