2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
5 * Copyright (c) 2001-2004, The GROMACS development team.
6 * Copyright (c) 2011-2019,2020,2021, by the GROMACS development team, led by
7 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
8 * and including many others, as listed in the AUTHORS file in the
9 * top-level source directory and at http://www.gromacs.org.
11 * GROMACS is free software; you can redistribute it and/or
12 * modify it under the terms of the GNU Lesser General Public License
13 * as published by the Free Software Foundation; either version 2.1
14 * of the License, or (at your option) any later version.
16 * GROMACS is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
19 * Lesser General Public License for more details.
21 * You should have received a copy of the GNU Lesser General Public
22 * License along with GROMACS; if not, see
23 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
24 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
26 * If you want to redistribute modifications to GROMACS, please
27 * consider that scientific software is very special. Version
28 * control is crucial - bugs must be traceable. We will be happy to
29 * consider code for inclusion in the official distribution, but
30 * derived work must not be called official GROMACS. Details are found
31 * in the README & COPYING files - if they are missing, get the
32 * official version at http://www.gromacs.org.
34 * To help us fund GROMACS development, we humbly ask that you cite
35 * the research papers on the package. Check out http://www.gromacs.org.
39 * \brief Implements the MD runner routine calling all integrators.
41 * \author David van der Spoel <david.vanderspoel@icm.uu.se>
42 * \ingroup module_mdrun
59 #include "gromacs/commandline/filenm.h"
60 #include "gromacs/domdec/builder.h"
61 #include "gromacs/domdec/domdec.h"
62 #include "gromacs/domdec/domdec_struct.h"
63 #include "gromacs/domdec/gpuhaloexchange.h"
64 #include "gromacs/domdec/localatomsetmanager.h"
65 #include "gromacs/domdec/partition.h"
66 #include "gromacs/domdec/reversetopology.h"
67 #include "gromacs/ewald/ewald_utils.h"
68 #include "gromacs/ewald/pme.h"
69 #include "gromacs/ewald/pme_gpu_program.h"
70 #include "gromacs/ewald/pme_only.h"
71 #include "gromacs/ewald/pme_pp_comm_gpu.h"
72 #include "gromacs/fileio/checkpoint.h"
73 #include "gromacs/fileio/gmxfio.h"
74 #include "gromacs/fileio/oenv.h"
75 #include "gromacs/fileio/tpxio.h"
76 #include "gromacs/gmxlib/network.h"
77 #include "gromacs/gmxlib/nrnb.h"
78 #include "gromacs/gpu_utils/device_stream_manager.h"
79 #include "gromacs/hardware/cpuinfo.h"
80 #include "gromacs/hardware/detecthardware.h"
81 #include "gromacs/hardware/device_management.h"
82 #include "gromacs/hardware/hardwaretopology.h"
83 #include "gromacs/hardware/printhardware.h"
84 #include "gromacs/imd/imd.h"
85 #include "gromacs/listed_forces/disre.h"
86 #include "gromacs/listed_forces/listed_forces_gpu.h"
87 #include "gromacs/listed_forces/listed_forces.h"
88 #include "gromacs/listed_forces/orires.h"
89 #include "gromacs/math/functions.h"
90 #include "gromacs/math/utilities.h"
91 #include "gromacs/math/vec.h"
92 #include "gromacs/mdlib/boxdeformation.h"
93 #include "gromacs/mdlib/broadcaststructs.h"
94 #include "gromacs/mdlib/calc_verletbuf.h"
95 #include "gromacs/mdlib/dispersioncorrection.h"
96 #include "gromacs/mdlib/enerdata_utils.h"
97 #include "gromacs/mdlib/force.h"
98 #include "gromacs/mdlib/forcerec.h"
99 #include "gromacs/mdlib/gmx_omp_nthreads.h"
100 #include "gromacs/mdlib/gpuforcereduction.h"
101 #include "gromacs/mdlib/makeconstraints.h"
102 #include "gromacs/mdlib/md_support.h"
103 #include "gromacs/mdlib/mdatoms.h"
104 #include "gromacs/mdlib/sighandler.h"
105 #include "gromacs/mdlib/stophandler.h"
106 #include "gromacs/mdlib/tgroup.h"
107 #include "gromacs/mdlib/updategroups.h"
108 #include "gromacs/mdlib/vsite.h"
109 #include "gromacs/mdrun/mdmodules.h"
110 #include "gromacs/mdrun/simulationcontext.h"
111 #include "gromacs/mdrun/simulationinput.h"
112 #include "gromacs/mdrun/simulationinputhandle.h"
113 #include "gromacs/mdrunutility/handlerestart.h"
114 #include "gromacs/mdrunutility/logging.h"
115 #include "gromacs/mdrunutility/multisim.h"
116 #include "gromacs/mdrunutility/printtime.h"
117 #include "gromacs/mdrunutility/threadaffinity.h"
118 #include "gromacs/mdtypes/checkpointdata.h"
119 #include "gromacs/mdtypes/commrec.h"
120 #include "gromacs/mdtypes/enerdata.h"
121 #include "gromacs/mdtypes/fcdata.h"
122 #include "gromacs/mdtypes/forcerec.h"
123 #include "gromacs/mdtypes/group.h"
124 #include "gromacs/mdtypes/inputrec.h"
125 #include "gromacs/mdtypes/interaction_const.h"
126 #include "gromacs/mdtypes/md_enums.h"
127 #include "gromacs/mdtypes/mdatom.h"
128 #include "gromacs/mdtypes/mdrunoptions.h"
129 #include "gromacs/mdtypes/observableshistory.h"
130 #include "gromacs/mdtypes/simulation_workload.h"
131 #include "gromacs/mdtypes/state.h"
132 #include "gromacs/mdtypes/state_propagator_data_gpu.h"
133 #include "gromacs/modularsimulator/modularsimulator.h"
134 #include "gromacs/nbnxm/gpu_data_mgmt.h"
135 #include "gromacs/nbnxm/nbnxm.h"
136 #include "gromacs/nbnxm/pairlist_tuning.h"
137 #include "gromacs/pbcutil/pbc.h"
138 #include "gromacs/pulling/output.h"
139 #include "gromacs/pulling/pull.h"
140 #include "gromacs/pulling/pull_rotation.h"
141 #include "gromacs/restraint/manager.h"
142 #include "gromacs/restraint/restraintmdmodule.h"
143 #include "gromacs/restraint/restraintpotential.h"
144 #include "gromacs/swap/swapcoords.h"
145 #include "gromacs/taskassignment/decidegpuusage.h"
146 #include "gromacs/taskassignment/decidesimulationworkload.h"
147 #include "gromacs/taskassignment/resourcedivision.h"
148 #include "gromacs/taskassignment/taskassignment.h"
149 #include "gromacs/taskassignment/usergpuids.h"
150 #include "gromacs/timing/gpu_timing.h"
151 #include "gromacs/timing/wallcycle.h"
152 #include "gromacs/timing/wallcyclereporting.h"
153 #include "gromacs/topology/mtop_util.h"
154 #include "gromacs/trajectory/trajectoryframe.h"
155 #include "gromacs/utility/basenetwork.h"
156 #include "gromacs/utility/cstringutil.h"
157 #include "gromacs/utility/exceptions.h"
158 #include "gromacs/utility/fatalerror.h"
159 #include "gromacs/utility/filestream.h"
160 #include "gromacs/utility/gmxassert.h"
161 #include "gromacs/utility/gmxmpi.h"
162 #include "gromacs/utility/keyvaluetree.h"
163 #include "gromacs/utility/logger.h"
164 #include "gromacs/utility/loggerbuilder.h"
165 #include "gromacs/utility/mdmodulesnotifiers.h"
166 #include "gromacs/utility/physicalnodecommunicator.h"
167 #include "gromacs/utility/pleasecite.h"
168 #include "gromacs/utility/programcontext.h"
169 #include "gromacs/utility/smalloc.h"
170 #include "gromacs/utility/stringutil.h"
171 #include "gromacs/utility/mpiinfo.h"
173 #include "isimulator.h"
174 #include "membedholder.h"
175 #include "replicaexchange.h"
176 #include "simulatorbuilder.h"
182 /*! \brief Manage any development feature flag variables encountered
184 * The use of dev features indicated by environment variables is
185 * logged in order to ensure that runs with such features enabled can
186 * be identified from their log and standard output. Any cross
187 * dependencies are also checked, and if unsatisfied, a fatal error
190 * Note that some development features overrides are applied already here:
191 * the GPU communication flags are set to false in non-tMPI and non-CUDA builds.
193 * \param[in] mdlog Logger object.
194 * \param[in] useGpuForNonbonded True if the nonbonded task is offloaded in this run.
195 * \param[in] pmeRunMode The PME run mode for this run
196 * \returns The object populated with development feature flags.
198 static DevelopmentFeatureFlags manageDevelopmentFeatures(const gmx::MDLogger& mdlog,
199 const bool useGpuForNonbonded,
200 const PmeRunMode pmeRunMode)
202 DevelopmentFeatureFlags devFlags;
204 // Some builds of GCC 5 give false positive warnings that these
205 // getenv results are ignored when clearly they are used.
206 #pragma GCC diagnostic push
207 #pragma GCC diagnostic ignored "-Wunused-result"
209 devFlags.enableGpuBufferOps =
210 GMX_GPU_CUDA && useGpuForNonbonded && (getenv("GMX_USE_GPU_BUFFER_OPS") != nullptr);
211 devFlags.enableGpuHaloExchange = GMX_GPU_CUDA && getenv("GMX_GPU_DD_COMMS") != nullptr;
212 devFlags.forceGpuUpdateDefault = (getenv("GMX_FORCE_UPDATE_DEFAULT_GPU") != nullptr) || GMX_FAHCORE;
213 devFlags.enableGpuPmePPComm = GMX_GPU_CUDA && getenv("GMX_GPU_PME_PP_COMMS") != nullptr;
215 #pragma GCC diagnostic pop
217 // Direct GPU comm path is being used with CUDA_AWARE_MPI
218 // make sure underlying MPI implementation is CUDA-aware
219 if (!GMX_THREAD_MPI && (devFlags.enableGpuPmePPComm || devFlags.enableGpuHaloExchange))
221 const bool haveDetectedCudaAwareMpi =
222 (checkMpiCudaAwareSupport() == CudaAwareMpiStatus::Supported);
223 const bool forceCudaAwareMpi = (getenv("GMX_FORCE_CUDA_AWARE_MPI") != nullptr);
225 if (!haveDetectedCudaAwareMpi && forceCudaAwareMpi)
227 // CUDA-aware support not detected in MPI library but, user has forced it's use
228 GMX_LOG(mdlog.warning)
230 .appendTextFormatted(
231 "This run has forced use of 'CUDA-aware MPI'. "
232 "But, GROMACS cannot determine if underlying MPI "
233 "is CUDA-aware. GROMACS recommends use of latest openMPI version "
234 "for CUDA-aware support. "
235 "If you observe failures at runtime, try unsetting "
236 "GMX_FORCE_CUDA_AWARE_MPI environment variable.");
239 if (haveDetectedCudaAwareMpi || forceCudaAwareMpi)
241 devFlags.usingCudaAwareMpi = true;
242 GMX_LOG(mdlog.warning)
244 .appendTextFormatted(
245 "Using CUDA-aware MPI for 'GPU halo exchange' or 'GPU PME-PP "
246 "communications' feature.");
250 if (devFlags.enableGpuHaloExchange)
252 GMX_LOG(mdlog.warning)
254 .appendTextFormatted(
255 "GMX_GPU_DD_COMMS environment variable detected, but the 'GPU "
256 "halo exchange' feature will not be enabled as GROMACS couldn't "
257 "detect CUDA_aware support in underlying MPI implementation.");
258 devFlags.enableGpuHaloExchange = false;
260 if (devFlags.enableGpuPmePPComm)
262 GMX_LOG(mdlog.warning)
265 "GMX_GPU_PME_PP_COMMS environment variable detected, but the "
266 "'GPU PME-PP communications' feature will not be enabled as "
268 "detect CUDA_aware support in underlying MPI implementation.");
269 devFlags.enableGpuPmePPComm = false;
272 GMX_LOG(mdlog.warning)
274 .appendTextFormatted(
275 "GROMACS recommends use of latest OpenMPI version for CUDA-aware "
277 "If you are certain about CUDA-aware support in your MPI library, "
278 "you can force it's use by setting environment variable "
279 " GMX_FORCE_CUDA_AWARE_MPI.");
283 if (devFlags.enableGpuBufferOps)
285 GMX_LOG(mdlog.warning)
287 .appendTextFormatted(
288 "This run uses the 'GPU buffer ops' feature, enabled by the "
289 "GMX_USE_GPU_BUFFER_OPS environment variable.");
292 if (devFlags.forceGpuUpdateDefault)
294 GMX_LOG(mdlog.warning)
296 .appendTextFormatted(
297 "This run will default to '-update gpu' as requested by the "
298 "GMX_FORCE_UPDATE_DEFAULT_GPU environment variable. GPU update with domain "
299 "decomposition lacks substantial testing and should be used with caution.");
302 if (devFlags.enableGpuHaloExchange)
304 if (useGpuForNonbonded)
306 if (!devFlags.enableGpuBufferOps)
308 GMX_LOG(mdlog.warning)
310 .appendTextFormatted(
311 "Enabling GPU buffer operations required by GMX_GPU_DD_COMMS "
312 "(equivalent with GMX_USE_GPU_BUFFER_OPS=1).");
313 devFlags.enableGpuBufferOps = true;
315 GMX_LOG(mdlog.warning)
317 .appendTextFormatted(
318 "This run has requested the 'GPU halo exchange' feature, enabled by "
320 "GMX_GPU_DD_COMMS environment variable.");
324 GMX_LOG(mdlog.warning)
326 .appendTextFormatted(
327 "GMX_GPU_DD_COMMS environment variable detected, but the 'GPU "
328 "halo exchange' feature will not be enabled as nonbonded interactions "
329 "are not offloaded.");
330 devFlags.enableGpuHaloExchange = false;
334 if (devFlags.enableGpuPmePPComm)
336 if (pmeRunMode == PmeRunMode::GPU)
338 if (!devFlags.enableGpuBufferOps)
340 GMX_LOG(mdlog.warning)
342 .appendTextFormatted(
343 "Enabling GPU buffer operations required by GMX_GPU_PME_PP_COMMS "
344 "(equivalent with GMX_USE_GPU_BUFFER_OPS=1).");
345 devFlags.enableGpuBufferOps = true;
347 GMX_LOG(mdlog.warning)
349 .appendTextFormatted(
350 "This run uses the 'GPU PME-PP communications' feature, enabled "
351 "by the GMX_GPU_PME_PP_COMMS environment variable.");
355 std::string clarification;
356 if (pmeRunMode == PmeRunMode::Mixed)
359 "PME FFT and gather are not offloaded to the GPU (PME is running in mixed "
364 clarification = "PME is not offloaded to the GPU.";
366 GMX_LOG(mdlog.warning)
369 "GMX_GPU_PME_PP_COMMS environment variable detected, but the "
370 "'GPU PME-PP communications' feature was not enabled as "
372 devFlags.enableGpuPmePPComm = false;
379 /*! \brief Barrier for safe simultaneous thread access to mdrunner data
381 * Used to ensure that the master thread does not modify mdrunner during copy
382 * on the spawned threads. */
383 static void threadMpiMdrunnerAccessBarrier()
386 MPI_Barrier(MPI_COMM_WORLD);
390 Mdrunner Mdrunner::cloneOnSpawnedThread() const
392 auto newRunner = Mdrunner(std::make_unique<MDModules>());
394 // All runners in the same process share a restraint manager resource because it is
395 // part of the interface to the client code, which is associated only with the
396 // original thread. Handles to the same resources can be obtained by copy.
398 newRunner.restraintManager_ = std::make_unique<RestraintManager>(*restraintManager_);
401 // Copy members of master runner.
402 // \todo Replace with builder when Simulation context and/or runner phases are better defined.
403 // Ref https://gitlab.com/gromacs/gromacs/-/issues/2587 and https://gitlab.com/gromacs/gromacs/-/issues/2375
404 newRunner.hw_opt = hw_opt;
405 newRunner.filenames = filenames;
407 newRunner.hwinfo_ = hwinfo_;
408 newRunner.oenv = oenv;
409 newRunner.mdrunOptions = mdrunOptions;
410 newRunner.domdecOptions = domdecOptions;
411 newRunner.nbpu_opt = nbpu_opt;
412 newRunner.pme_opt = pme_opt;
413 newRunner.pme_fft_opt = pme_fft_opt;
414 newRunner.bonded_opt = bonded_opt;
415 newRunner.update_opt = update_opt;
416 newRunner.nstlist_cmdline = nstlist_cmdline;
417 newRunner.replExParams = replExParams;
418 newRunner.pforce = pforce;
419 // Give the spawned thread the newly created valid communicator
420 // for the simulation.
421 newRunner.libraryWorldCommunicator = MPI_COMM_WORLD;
422 newRunner.simulationCommunicator = MPI_COMM_WORLD;
424 newRunner.startingBehavior = startingBehavior;
425 newRunner.stopHandlerBuilder_ = std::make_unique<StopHandlerBuilder>(*stopHandlerBuilder_);
426 newRunner.inputHolder_ = inputHolder_;
428 threadMpiMdrunnerAccessBarrier();
433 /*! \brief The callback used for running on spawned threads.
435 * Obtains the pointer to the master mdrunner object from the one
436 * argument permitted to the thread-launch API call, copies it to make
437 * a new runner for this thread, reinitializes necessary data, and
438 * proceeds to the simulation. */
439 static void mdrunner_start_fn(const void* arg)
443 const auto* masterMdrunner = reinterpret_cast<const gmx::Mdrunner*>(arg);
444 /* copy the arg list to make sure that it's thread-local. This
445 doesn't copy pointed-to items, of course; fnm, cr and fplog
446 are reset in the call below, all others should be const. */
447 gmx::Mdrunner mdrunner = masterMdrunner->cloneOnSpawnedThread();
450 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
454 void Mdrunner::spawnThreads(int numThreadsToLaunch)
457 /* now spawn new threads that start mdrunner_start_fn(), while
458 the main thread returns. Thread affinity is handled later. */
459 if (tMPI_Init_fn(TRUE, numThreadsToLaunch, TMPI_AFFINITY_NONE, mdrunner_start_fn, static_cast<const void*>(this))
462 GMX_THROW(gmx::InternalError("Failed to spawn thread-MPI threads"));
465 // Give the master thread the newly created valid communicator for
467 libraryWorldCommunicator = MPI_COMM_WORLD;
468 simulationCommunicator = MPI_COMM_WORLD;
469 threadMpiMdrunnerAccessBarrier();
471 GMX_UNUSED_VALUE(numThreadsToLaunch);
472 GMX_UNUSED_VALUE(mdrunner_start_fn);
478 /*! \brief Initialize variables for Verlet scheme simulation */
479 static void prepare_verlet_scheme(FILE* fplog,
483 const gmx_mtop_t& mtop,
485 bool makeGpuPairList,
486 const gmx::CpuInfo& cpuinfo)
488 // We checked the cut-offs in grompp, but double-check here.
489 // We have PME+LJcutoff kernels for rcoulomb>rvdw.
490 if (EEL_PME_EWALD(ir->coulombtype) && ir->vdwtype == VanDerWaalsType::Cut)
492 GMX_RELEASE_ASSERT(ir->rcoulomb >= ir->rvdw,
493 "With Verlet lists and PME we should have rcoulomb>=rvdw");
497 GMX_RELEASE_ASSERT(ir->rcoulomb == ir->rvdw,
498 "With Verlet lists and no PME rcoulomb and rvdw should be identical");
500 /* For NVE simulations, we will retain the initial list buffer */
501 if (EI_DYNAMICS(ir->eI) && ir->verletbuf_tol > 0
502 && !(EI_MD(ir->eI) && ir->etc == TemperatureCoupling::No))
504 /* Update the Verlet buffer size for the current run setup */
506 /* Here we assume SIMD-enabled kernels are being used. But as currently
507 * calc_verlet_buffer_size gives the same results for 4x8 and 4x4
508 * and 4x2 gives a larger buffer than 4x4, this is ok.
510 ListSetupType listType =
511 (makeGpuPairList ? ListSetupType::Gpu : ListSetupType::CpuSimdWhenSupported);
512 VerletbufListSetup listSetup = verletbufGetSafeListSetup(listType);
514 const real rlist_new =
515 calcVerletBufferSize(mtop, det(box), *ir, ir->nstlist, ir->nstlist - 1, -1, listSetup);
517 if (rlist_new != ir->rlist)
519 if (fplog != nullptr)
522 "\nChanging rlist from %g to %g for non-bonded %dx%d atom kernels\n\n",
525 listSetup.cluster_size_i,
526 listSetup.cluster_size_j);
528 ir->rlist = rlist_new;
532 if (nstlist_cmdline > 0 && (!EI_DYNAMICS(ir->eI) || ir->verletbuf_tol <= 0))
535 "Can not set nstlist without %s",
536 !EI_DYNAMICS(ir->eI) ? "dynamics" : "verlet-buffer-tolerance");
539 if (EI_DYNAMICS(ir->eI))
541 /* Set or try nstlist values */
542 increaseNstlist(fplog, cr, ir, nstlist_cmdline, &mtop, box, makeGpuPairList, cpuinfo);
546 /*! \brief Override the nslist value in inputrec
548 * with value passed on the command line (if any)
550 static void override_nsteps_cmdline(const gmx::MDLogger& mdlog, int64_t nsteps_cmdline, t_inputrec* ir)
554 /* override with anything else than the default -2 */
555 if (nsteps_cmdline > -2)
557 char sbuf_steps[STEPSTRSIZE];
558 char sbuf_msg[STRLEN];
560 ir->nsteps = nsteps_cmdline;
561 if (EI_DYNAMICS(ir->eI) && nsteps_cmdline != -1)
564 "Overriding nsteps with value passed on the command line: %s steps, %.3g ps",
565 gmx_step_str(nsteps_cmdline, sbuf_steps),
566 fabs(nsteps_cmdline * ir->delta_t));
571 "Overriding nsteps with value passed on the command line: %s steps",
572 gmx_step_str(nsteps_cmdline, sbuf_steps));
575 GMX_LOG(mdlog.warning).asParagraph().appendText(sbuf_msg);
577 else if (nsteps_cmdline < -2)
579 gmx_fatal(FARGS, "Invalid nsteps value passed on the command line: %" PRId64, nsteps_cmdline);
581 /* Do nothing if nsteps_cmdline == -2 */
587 /*! \brief Return whether GPU acceleration of nonbondeds is supported with the given settings.
589 * If not, and if a warning may be issued, logs a warning about
590 * falling back to CPU code. With thread-MPI, only the first
591 * call to this function should have \c issueWarning true. */
592 static bool gpuAccelerationOfNonbondedIsUseful(const MDLogger& mdlog, const t_inputrec& ir, bool issueWarning)
594 bool gpuIsUseful = true;
597 if (ir.opts.ngener - ir.nwall > 1)
599 /* The GPU code does not support more than one energy group.
600 * If the user requested GPUs explicitly, a fatal error is given later.
604 "Multiple energy groups is not implemented for GPUs, falling back to the CPU. "
605 "For better performance, run on the GPU without energy groups and then do "
606 "gmx mdrun -rerun option on the trajectory with an energy group .tpr file.";
612 warning = "TPI is not implemented for GPUs.";
615 if (!gpuIsUseful && issueWarning)
617 GMX_LOG(mdlog.warning).asParagraph().appendText(warning);
623 //! Initializes the logger for mdrun.
624 static gmx::LoggerOwner buildLogger(FILE* fplog, const bool isSimulationMasterRank)
626 gmx::LoggerBuilder builder;
627 if (fplog != nullptr)
629 builder.addTargetFile(gmx::MDLogger::LogLevel::Info, fplog);
631 if (isSimulationMasterRank)
633 builder.addTargetStream(gmx::MDLogger::LogLevel::Warning, &gmx::TextOutputFile::standardError());
635 return builder.build();
638 //! Make a TaskTarget from an mdrun argument string.
639 static TaskTarget findTaskTarget(const char* optionString)
641 TaskTarget returnValue = TaskTarget::Auto;
643 if (strncmp(optionString, "auto", 3) == 0)
645 returnValue = TaskTarget::Auto;
647 else if (strncmp(optionString, "cpu", 3) == 0)
649 returnValue = TaskTarget::Cpu;
651 else if (strncmp(optionString, "gpu", 3) == 0)
653 returnValue = TaskTarget::Gpu;
657 GMX_ASSERT(false, "Option string should have been checked for sanity already");
663 //! Finish run, aggregate data to print performance info.
664 static void finish_run(FILE* fplog,
665 const gmx::MDLogger& mdlog,
667 const t_inputrec& inputrec,
669 gmx_wallcycle* wcycle,
670 gmx_walltime_accounting_t walltime_accounting,
671 nonbonded_verlet_t* nbv,
672 const gmx_pme_t* pme,
676 double nbfs = 0, mflop = 0;
677 double elapsed_time, elapsed_time_over_all_ranks, elapsed_time_over_all_threads,
678 elapsed_time_over_all_threads_over_all_ranks;
679 /* Control whether it is valid to print a report. Only the
680 simulation master may print, but it should not do so if the run
681 terminated e.g. before a scheduled reset step. This is
682 complicated by the fact that PME ranks are unaware of the
683 reason why they were sent a pmerecvqxFINISH. To avoid
684 communication deadlocks, we always do the communication for the
685 report, even if we've decided not to write the report, because
686 how long it takes to finish the run is not important when we've
687 decided not to report on the simulation performance.
689 Further, we only report performance for dynamical integrators,
690 because those are the only ones for which we plan to
691 consider doing any optimizations. */
692 bool printReport = EI_DYNAMICS(inputrec.eI) && SIMMASTER(cr);
694 if (printReport && !walltime_accounting_get_valid_finish(walltime_accounting))
696 GMX_LOG(mdlog.warning)
698 .appendText("Simulation ended prematurely, no performance report will be written.");
703 std::unique_ptr<t_nrnb> nrnbTotalStorage;
706 nrnbTotalStorage = std::make_unique<t_nrnb>();
707 nrnb_tot = nrnbTotalStorage.get();
709 MPI_Allreduce(nrnb->n.data(), nrnb_tot->n.data(), eNRNB, MPI_DOUBLE, MPI_SUM, cr->mpi_comm_mysim);
717 elapsed_time = walltime_accounting_get_time_since_reset(walltime_accounting);
718 elapsed_time_over_all_threads =
719 walltime_accounting_get_time_since_reset_over_all_threads(walltime_accounting);
723 /* reduce elapsed_time over all MPI ranks in the current simulation */
724 MPI_Allreduce(&elapsed_time, &elapsed_time_over_all_ranks, 1, MPI_DOUBLE, MPI_SUM, cr->mpi_comm_mysim);
725 elapsed_time_over_all_ranks /= cr->nnodes;
726 /* Reduce elapsed_time_over_all_threads over all MPI ranks in the
727 * current simulation. */
728 MPI_Allreduce(&elapsed_time_over_all_threads,
729 &elapsed_time_over_all_threads_over_all_ranks,
738 elapsed_time_over_all_ranks = elapsed_time;
739 elapsed_time_over_all_threads_over_all_ranks = elapsed_time_over_all_threads;
744 print_flop(fplog, nrnb_tot, &nbfs, &mflop);
747 if (thisRankHasDuty(cr, DUTY_PP) && DOMAINDECOMP(cr))
749 print_dd_statistics(cr, inputrec, fplog);
752 /* TODO Move the responsibility for any scaling by thread counts
753 * to the code that handled the thread region, so that there's a
754 * mechanism to keep cycle counting working during the transition
755 * to task parallelism. */
756 int nthreads_pp = gmx_omp_nthreads_get(ModuleMultiThread::Nonbonded);
757 int nthreads_pme = gmx_omp_nthreads_get(ModuleMultiThread::Pme);
758 wallcycle_scale_by_num_threads(
759 wcycle, thisRankHasDuty(cr, DUTY_PME) && !thisRankHasDuty(cr, DUTY_PP), nthreads_pp, nthreads_pme);
760 auto cycle_sum(wallcycle_sum(cr, wcycle));
764 auto* nbnxn_gpu_timings =
765 (nbv != nullptr && nbv->useGpu()) ? Nbnxm::gpu_get_timings(nbv->gpu_nbv) : nullptr;
766 gmx_wallclock_gpu_pme_t pme_gpu_timings = {};
768 if (pme_gpu_task_enabled(pme))
770 pme_gpu_get_timings(pme, &pme_gpu_timings);
772 wallcycle_print(fplog,
778 elapsed_time_over_all_ranks,
784 if (EI_DYNAMICS(inputrec.eI))
786 delta_t = inputrec.delta_t;
792 elapsed_time_over_all_threads_over_all_ranks,
793 elapsed_time_over_all_ranks,
794 walltime_accounting_get_nsteps_done_since_reset(walltime_accounting),
802 elapsed_time_over_all_threads_over_all_ranks,
803 elapsed_time_over_all_ranks,
804 walltime_accounting_get_nsteps_done_since_reset(walltime_accounting),
812 int Mdrunner::mdrunner()
815 std::unique_ptr<t_forcerec> fr;
816 real ewaldcoeff_q = 0;
817 real ewaldcoeff_lj = 0;
818 int nChargePerturbed = -1, nTypePerturbed = 0;
819 gmx_walltime_accounting_t walltime_accounting = nullptr;
820 MembedHolder membedHolder(filenames.size(), filenames.data());
822 /* CAUTION: threads may be started later on in this function, so
823 cr doesn't reflect the final parallel state right now */
826 /* TODO: inputrec should tell us whether we use an algorithm, not a file option */
827 const bool doEssentialDynamics = opt2bSet("-ei", filenames.size(), filenames.data());
828 const bool doRerun = mdrunOptions.rerun;
830 // Handle task-assignment related user options.
831 EmulateGpuNonbonded emulateGpuNonbonded =
832 (getenv("GMX_EMULATE_GPU") != nullptr ? EmulateGpuNonbonded::Yes : EmulateGpuNonbonded::No);
834 std::vector<int> userGpuTaskAssignment;
837 userGpuTaskAssignment = parseUserTaskAssignmentString(hw_opt.userGpuTaskAssignment);
839 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
840 auto nonbondedTarget = findTaskTarget(nbpu_opt);
841 auto pmeTarget = findTaskTarget(pme_opt);
842 auto pmeFftTarget = findTaskTarget(pme_fft_opt);
843 auto bondedTarget = findTaskTarget(bonded_opt);
844 auto updateTarget = findTaskTarget(update_opt);
846 FILE* fplog = nullptr;
847 // If we are appending, we don't write log output because we need
848 // to check that the old log file matches what the checkpoint file
849 // expects. Otherwise, we should start to write log output now if
850 // there is a file ready for it.
851 if (logFileHandle != nullptr && startingBehavior != StartingBehavior::RestartWithAppending)
853 fplog = gmx_fio_getfp(logFileHandle);
855 const bool isSimulationMasterRank = findIsSimulationMasterRank(ms, simulationCommunicator);
856 gmx::LoggerOwner logOwner(buildLogger(fplog, isSimulationMasterRank));
857 gmx::MDLogger mdlog(logOwner.logger());
859 gmx_print_detected_hardware(fplog, isSimulationMasterRank && isMasterSim(ms), mdlog, hwinfo_);
861 std::vector<int> availableDevices =
862 makeListOfAvailableDevices(hwinfo_->deviceInfoList, hw_opt.devicesSelectedByUser);
863 const int numAvailableDevices = gmx::ssize(availableDevices);
865 // Print citation requests after all software/hardware printing
866 pleaseCiteGromacs(fplog);
868 // Note: legacy program logic relies on checking whether these pointers are assigned.
869 // Objects may or may not be allocated later.
870 std::unique_ptr<t_inputrec> inputrec;
871 std::unique_ptr<t_state> globalState;
873 auto partialDeserializedTpr = std::make_unique<PartialDeserializedTprFile>();
875 if (isSimulationMasterRank)
877 // Allocate objects to be initialized by later function calls.
878 /* Only the master rank has the global state */
879 globalState = std::make_unique<t_state>();
880 inputrec = std::make_unique<t_inputrec>();
882 /* Read (nearly) all data required for the simulation
883 * and keep the partly serialized tpr contents to send to other ranks later
885 applyGlobalSimulationState(
886 *inputHolder_.get(), partialDeserializedTpr.get(), globalState.get(), inputrec.get(), &mtop);
889 /* Check and update the hardware options for internal consistency */
890 checkAndUpdateHardwareOptions(
891 mdlog, &hw_opt, isSimulationMasterRank, domdecOptions.numPmeRanks, inputrec.get());
893 if (GMX_THREAD_MPI && isSimulationMasterRank)
895 bool useGpuForNonbonded = false;
896 bool useGpuForPme = false;
899 GMX_RELEASE_ASSERT(inputrec != nullptr, "Keep the compiler happy");
901 // If the user specified the number of ranks, then we must
902 // respect that, but in default mode, we need to allow for
903 // the number of GPUs to choose the number of ranks.
904 auto canUseGpuForNonbonded = buildSupportsNonbondedOnGpu(nullptr);
905 useGpuForNonbonded = decideWhetherToUseGpusForNonbondedWithThreadMpi(
907 numAvailableDevices > 0,
908 userGpuTaskAssignment,
910 canUseGpuForNonbonded,
911 gpuAccelerationOfNonbondedIsUseful(mdlog, *inputrec, GMX_THREAD_MPI),
912 hw_opt.nthreads_tmpi);
913 useGpuForPme = decideWhetherToUseGpusForPmeWithThreadMpi(useGpuForNonbonded,
916 userGpuTaskAssignment,
919 hw_opt.nthreads_tmpi,
920 domdecOptions.numPmeRanks);
922 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
924 /* Determine how many thread-MPI ranks to start.
926 * TODO Over-writing the user-supplied value here does
927 * prevent any possible subsequent checks from working
929 hw_opt.nthreads_tmpi = get_nthreads_mpi(hwinfo_,
937 membedHolder.doMembed());
939 // Now start the threads for thread MPI.
940 spawnThreads(hw_opt.nthreads_tmpi);
941 // The spawned threads enter mdrunner() and execution of
942 // master and spawned threads joins at the end of this block.
945 GMX_RELEASE_ASSERT(ms || simulationCommunicator != MPI_COMM_NULL,
946 "Must have valid communicator unless running a multi-simulation");
947 CommrecHandle crHandle = init_commrec(simulationCommunicator);
948 t_commrec* cr = crHandle.get();
949 GMX_RELEASE_ASSERT(cr != nullptr, "Must have valid commrec");
951 PhysicalNodeCommunicator physicalNodeComm(libraryWorldCommunicator, gmx_physicalnode_id_hash());
953 // If we detected the topology on this system, double-check that it makes sense
954 if (hwinfo_->hardwareTopology->isThisSystem())
956 hardwareTopologyDoubleCheckDetection(mdlog, *hwinfo_->hardwareTopology);
961 /* now broadcast everything to the non-master nodes/threads: */
962 if (!isSimulationMasterRank)
964 // Until now, only the master rank has a non-null pointer.
965 // On non-master ranks, allocate the object that will receive data in the following call.
966 inputrec = std::make_unique<t_inputrec>();
968 init_parallel(cr->mpiDefaultCommunicator,
972 partialDeserializedTpr.get());
974 GMX_RELEASE_ASSERT(inputrec != nullptr, "All ranks should have a valid inputrec now");
975 partialDeserializedTpr.reset(nullptr);
978 !inputrec->useConstantAcceleration,
979 "Linear acceleration has been removed in GROMACS 2022, and was broken for many years "
980 "before that. Use GROMACS 4.5 or earlier if you need this feature.");
982 // Now the number of ranks is known to all ranks, and each knows
983 // the inputrec read by the master rank. The ranks can now all run
984 // the task-deciding functions and will agree on the result
985 // without needing to communicate.
986 const bool useDomainDecomposition =
987 (PAR(cr) && !(EI_TPI(inputrec->eI) || inputrec->eI == IntegrationAlgorithm::NM));
989 // Note that these variables describe only their own node.
991 // Note that when bonded interactions run on a GPU they always run
992 // alongside a nonbonded task, so do not influence task assignment
993 // even though they affect the force calculation workload.
994 bool useGpuForNonbonded = false;
995 bool useGpuForPme = false;
996 bool useGpuForBonded = false;
997 bool useGpuForUpdate = false;
998 bool gpusWereDetected = hwinfo_->ngpu_compatible_tot > 0;
1001 // It's possible that there are different numbers of GPUs on
1002 // different nodes, which is the user's responsibility to
1003 // handle. If unsuitable, we will notice that during task
1005 auto canUseGpuForNonbonded = buildSupportsNonbondedOnGpu(nullptr);
1006 useGpuForNonbonded = decideWhetherToUseGpusForNonbonded(
1008 userGpuTaskAssignment,
1009 emulateGpuNonbonded,
1010 canUseGpuForNonbonded,
1011 gpuAccelerationOfNonbondedIsUseful(mdlog, *inputrec, !GMX_THREAD_MPI),
1013 useGpuForPme = decideWhetherToUseGpusForPme(useGpuForNonbonded,
1015 userGpuTaskAssignment,
1018 cr->sizeOfDefaultCommunicator,
1019 domdecOptions.numPmeRanks,
1021 useGpuForBonded = decideWhetherToUseGpusForBonded(
1022 useGpuForNonbonded, useGpuForPme, bondedTarget, *inputrec, mtop, domdecOptions.numPmeRanks, gpusWereDetected);
1024 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1026 const PmeRunMode pmeRunMode = determinePmeRunMode(useGpuForPme, pmeFftTarget, *inputrec);
1028 // Initialize development feature flags that enabled by environment variable
1029 // and report those features that are enabled.
1030 const DevelopmentFeatureFlags devFlags =
1031 manageDevelopmentFeatures(mdlog, useGpuForNonbonded, pmeRunMode);
1033 const bool useModularSimulator = checkUseModularSimulator(false,
1040 doEssentialDynamics,
1041 membedHolder.doMembed());
1043 // Build restraints.
1044 // TODO: hide restraint implementation details from Mdrunner.
1045 // There is nothing unique about restraints at this point as far as the
1046 // Mdrunner is concerned. The Mdrunner should just be getting a sequence of
1047 // factory functions from the SimulationContext on which to call mdModules_->add().
1048 // TODO: capture all restraints into a single RestraintModule, passed to the runner builder.
1049 for (auto&& restraint : restraintManager_->getRestraints())
1051 auto module = RestraintMDModule::create(restraint, restraint->sites());
1052 mdModules_->add(std::move(module));
1055 // TODO: Error handling
1056 mdModules_->assignOptionsToModules(*inputrec->params, nullptr);
1057 // now that the MDModules know their options, they know which callbacks to sign up to
1058 mdModules_->subscribeToSimulationSetupNotifications();
1059 const auto& setupNotifier = mdModules_->notifiers().simulationSetupNotifier_;
1061 if (inputrec->internalParameters != nullptr)
1063 setupNotifier.notify(*inputrec->internalParameters);
1066 if (fplog != nullptr)
1068 pr_inputrec(fplog, 0, "Input Parameters", inputrec.get(), FALSE);
1069 fprintf(fplog, "\n");
1074 /* In rerun, set velocities to zero if present */
1075 if (doRerun && ((globalState->flags & enumValueToBitMask(StateEntry::V)) != 0))
1077 // rerun does not use velocities
1081 "Rerun trajectory contains velocities. Rerun does only evaluate "
1082 "potential energy and forces. The velocities will be ignored.");
1083 for (int i = 0; i < globalState->natoms; i++)
1085 clear_rvec(globalState->v[i]);
1087 globalState->flags &= ~enumValueToBitMask(StateEntry::V);
1090 /* now make sure the state is initialized and propagated */
1091 set_state_entries(globalState.get(), inputrec.get(), useModularSimulator);
1094 /* NM and TPI parallelize over force/energy calculations, not atoms,
1095 * so we need to initialize and broadcast the global state.
1097 if (inputrec->eI == IntegrationAlgorithm::NM || inputrec->eI == IntegrationAlgorithm::TPI)
1101 globalState = std::make_unique<t_state>();
1103 broadcastStateWithoutDynamics(
1104 cr->mpiDefaultCommunicator, DOMAINDECOMP(cr), PAR(cr), globalState.get());
1107 /* A parallel command line option consistency check that we can
1108 only do after any threads have started. */
1110 && (domdecOptions.numCells[XX] > 1 || domdecOptions.numCells[YY] > 1
1111 || domdecOptions.numCells[ZZ] > 1 || domdecOptions.numPmeRanks > 0))
1114 "The -dd or -npme option request a parallel simulation, "
1116 "but %s was compiled without threads or MPI enabled",
1117 output_env_get_program_display_name(oenv));
1118 #elif GMX_THREAD_MPI
1119 "but the number of MPI-threads (option -ntmpi) is not set or is 1");
1121 "but %s was not started through mpirun/mpiexec or only one rank was requested "
1122 "through mpirun/mpiexec",
1123 output_env_get_program_display_name(oenv));
1127 if (doRerun && (EI_ENERGY_MINIMIZATION(inputrec->eI) || IntegrationAlgorithm::NM == inputrec->eI))
1130 "The .mdp file specified an energy mininization or normal mode algorithm, and "
1131 "these are not compatible with mdrun -rerun");
1134 /* NMR restraints must be initialized before load_checkpoint,
1135 * since with time averaging the history is added to t_state.
1136 * For proper consistency check we therefore need to extend
1138 * So the PME-only nodes (if present) will also initialize
1139 * the distance restraints.
1142 /* This needs to be called before read_checkpoint to extend the state */
1143 t_disresdata* disresdata;
1144 snew(disresdata, 1);
1148 DisResRunMode::MDRun,
1149 MASTER(cr) ? DDRole::Master : DDRole::Agent,
1150 PAR(cr) ? NumRanks::Multiple : NumRanks::Single,
1155 replExParams.exchangeInterval > 0);
1157 std::unique_ptr<t_oriresdata> oriresData;
1158 if (gmx_mtop_ftype_count(mtop, F_ORIRES) > 0)
1160 oriresData = std::make_unique<t_oriresdata>(fplog, mtop, *inputrec, cr, ms, globalState.get());
1163 auto deform = prepareBoxDeformation(globalState != nullptr ? globalState->box : box,
1164 MASTER(cr) ? DDRole::Master : DDRole::Agent,
1165 PAR(cr) ? NumRanks::Multiple : NumRanks::Single,
1166 cr->mpi_comm_mygroup,
1170 /* We have to remember the generation's first step before reading checkpoint.
1171 This way, we can report to the F@H core both the generation's first step
1172 and the restored first step, thus making it able to distinguish between
1173 an interruption/resume and start of the n-th generation simulation.
1174 Having this information, the F@H core can correctly calculate and report
1177 int gen_first_step = 0;
1180 gen_first_step = inputrec->init_step;
1184 ObservablesHistory observablesHistory = {};
1186 auto modularSimulatorCheckpointData = std::make_unique<ReadCheckpointDataHolder>();
1187 if (startingBehavior != StartingBehavior::NewSimulation)
1189 /* Check if checkpoint file exists before doing continuation.
1190 * This way we can use identical input options for the first and subsequent runs...
1192 if (mdrunOptions.numStepsCommandline > -2)
1194 /* Temporarily set the number of steps to unlimited to avoid
1195 * triggering the nsteps check in load_checkpoint().
1196 * This hack will go away soon when the -nsteps option is removed.
1198 inputrec->nsteps = -1;
1201 // Finish applying initial simulation state information from external sources on all ranks.
1202 // Reconcile checkpoint file data with Mdrunner state established up to this point.
1203 applyLocalState(*inputHolder_.get(),
1206 domdecOptions.numCells,
1209 &observablesHistory,
1210 mdrunOptions.reproducible,
1211 mdModules_->notifiers(),
1212 modularSimulatorCheckpointData.get(),
1213 useModularSimulator);
1214 // TODO: (#3652) Synchronize filesystem state, SimulationInput contents, and program
1216 // on all code paths.
1217 // Write checkpoint or provide hook to update SimulationInput.
1218 // If there was a checkpoint file, SimulationInput contains more information
1219 // than if there wasn't. At this point, we have synchronized the in-memory
1220 // state with the filesystem state only for restarted simulations. We should
1221 // be calling applyLocalState unconditionally and expect that the completeness
1222 // of SimulationInput is not dependent on its creation method.
1224 if (startingBehavior == StartingBehavior::RestartWithAppending && logFileHandle)
1226 // Now we can start normal logging to the truncated log file.
1227 fplog = gmx_fio_getfp(logFileHandle);
1228 prepareLogAppending(fplog);
1229 logOwner = buildLogger(fplog, MASTER(cr));
1230 mdlog = logOwner.logger();
1237 fcRegisterSteps(inputrec->nsteps + inputrec->init_step, gen_first_step);
1241 if (mdrunOptions.numStepsCommandline > -2)
1246 "The -nsteps functionality is deprecated, and may be removed in a future "
1248 "Consider using gmx convert-tpr -nsteps or changing the appropriate .mdp "
1251 /* override nsteps with value set on the commandline */
1252 override_nsteps_cmdline(mdlog, mdrunOptions.numStepsCommandline, inputrec.get());
1254 if (isSimulationMasterRank)
1256 copy_mat(globalState->box, box);
1261 gmx_bcast(sizeof(box), box, cr->mpiDefaultCommunicator);
1264 if (inputrec->cutoff_scheme != CutoffScheme::Verlet)
1267 "This group-scheme .tpr file can no longer be run by mdrun. Please update to the "
1268 "Verlet scheme, or use an earlier version of GROMACS if necessary.");
1270 /* Update rlist and nstlist. */
1271 /* Note: prepare_verlet_scheme is calling increaseNstlist(...), which (while attempting to
1272 * increase rlist) tries to check if the newly chosen value fits with the DD scheme. As this is
1273 * run before any DD scheme is set up, this check is never executed. See #3334 for more details.
1275 prepare_verlet_scheme(fplog,
1281 useGpuForNonbonded || (emulateGpuNonbonded == EmulateGpuNonbonded::Yes),
1284 // We need to decide on update groups early, as this affects
1285 // inter-domain communication distances.
1286 auto updateGroupingsPerMoleculeType = makeUpdateGroupingsPerMoleculeType(mtop);
1287 const real maxUpdateGroupRadius = computeMaxUpdateGroupRadius(
1288 mtop, updateGroupingsPerMoleculeType, maxReferenceTemperature(*inputrec));
1289 const real cutoffMargin = std::sqrt(max_cutoff2(inputrec->pbcType, box)) - inputrec->rlist;
1290 UpdateGroups updateGroups = makeUpdateGroups(mdlog,
1291 std::move(updateGroupingsPerMoleculeType),
1292 maxUpdateGroupRadius,
1293 useDomainDecomposition,
1294 systemHasConstraintsOrVsites(mtop),
1297 // This builder is necessary while we have multi-part construction
1298 // of DD. Before DD is constructed, we use the existence of
1299 // the builder object to indicate that further construction of DD
1301 std::unique_ptr<DomainDecompositionBuilder> ddBuilder;
1302 if (useDomainDecomposition)
1304 ddBuilder = std::make_unique<DomainDecompositionBuilder>(
1311 mdModules_->notifiers(),
1313 updateGroups.updateGroupingPerMoleculeType(),
1314 updateGroups.useUpdateGroups(),
1315 updateGroups.maxUpdateGroupRadius(),
1316 positionsFromStatePointer(globalState.get()),
1322 /* PME, if used, is done on all nodes with 1D decomposition */
1323 cr->nnodes = cr->sizeOfDefaultCommunicator;
1324 cr->sim_nodeid = cr->rankInDefaultCommunicator;
1325 cr->nodeid = cr->rankInDefaultCommunicator;
1327 cr->duty = (DUTY_PP | DUTY_PME);
1329 if (inputrec->pbcType == PbcType::Screw)
1331 gmx_fatal(FARGS, "pbc=screw is only implemented with domain decomposition");
1335 // Produce the task assignment for this rank - done after DD is constructed
1336 GpuTaskAssignments gpuTaskAssignments = GpuTaskAssignmentsBuilder::build(
1338 userGpuTaskAssignment,
1340 simulationCommunicator,
1348 thisRankHasDuty(cr, DUTY_PP),
1349 // TODO cr->duty & DUTY_PME should imply that a PME
1350 // algorithm is active, but currently does not.
1351 EEL_PME(inputrec->coulombtype) && thisRankHasDuty(cr, DUTY_PME));
1353 // Get the device handles for the modules, nullptr when no task is assigned.
1355 DeviceInformation* deviceInfo = gpuTaskAssignments.initDevice(&deviceId);
1357 // timing enabling - TODO put this in gpu_utils (even though generally this is just option handling?)
1358 bool useTiming = true;
1362 /* WARNING: CUDA timings are incorrect with multiple streams.
1363 * This is the main reason why they are disabled by default.
1365 // TODO: Consider turning on by default when we can detect nr of streams.
1366 useTiming = (getenv("GMX_ENABLE_GPU_TIMING") != nullptr);
1368 else if (GMX_GPU_OPENCL)
1370 useTiming = (getenv("GMX_DISABLE_GPU_TIMING") == nullptr);
1373 // TODO Currently this is always built, yet DD partition code
1374 // checks if it is built before using it. Probably it should
1375 // become an MDModule that is made only when another module
1376 // requires it (e.g. pull, CompEl, density fitting), so that we
1377 // don't update the local atom sets unilaterally every step.
1378 LocalAtomSetManager atomSets;
1381 // TODO Pass the GPU streams to ddBuilder to use in buffer
1382 // transfers (e.g. halo exchange)
1383 cr->dd = ddBuilder->build(&atomSets);
1384 // The builder's job is done, so destruct it
1385 ddBuilder.reset(nullptr);
1386 // Note that local state still does not exist yet.
1389 // The GPU update is decided here because we need to know whether the constraints or
1390 // SETTLEs can span accross the domain borders (i.e. whether or not update groups are
1391 // defined). This is only known after DD is initialized, hence decision on using GPU
1392 // update is done so late.
1395 const bool haveFrozenAtoms = inputrecFrozenAtoms(inputrec.get());
1397 useGpuForUpdate = decideWhetherToUseGpuForUpdate(useDomainDecomposition,
1398 updateGroups.useUpdateGroups(),
1400 domdecOptions.numPmeRanks > 0,
1406 doEssentialDynamics,
1407 gmx_mtop_ftype_count(mtop, F_ORIRES) > 0,
1413 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1415 const bool printHostName = (cr->nnodes > 1);
1416 gpuTaskAssignments.reportGpuUsage(mdlog, printHostName, useGpuForBonded, pmeRunMode, useGpuForUpdate);
1418 const bool disableNonbondedCalculation = (getenv("GMX_NO_NONBONDED") != nullptr);
1419 if (disableNonbondedCalculation)
1421 /* turn off non-bonded calculations */
1422 GMX_LOG(mdlog.warning)
1425 "Found environment variable GMX_NO_NONBONDED.\n"
1426 "Disabling nonbonded calculations.");
1429 MdrunScheduleWorkload runScheduleWork;
1431 bool useGpuDirectHalo = decideWhetherToUseGpuForHalo(devFlags,
1432 havePPDomainDecomposition(cr),
1434 useModularSimulator,
1436 EI_ENERGY_MINIMIZATION(inputrec->eI));
1438 // Also populates the simulation constant workload description.
1439 runScheduleWork.simulationWork = createSimulationWorkload(*inputrec,
1440 disableNonbondedCalculation,
1448 std::unique_ptr<DeviceStreamManager> deviceStreamManager = nullptr;
1450 if (deviceInfo != nullptr)
1452 if (DOMAINDECOMP(cr) && thisRankHasDuty(cr, DUTY_PP))
1454 dd_setup_dlb_resource_sharing(cr, deviceId);
1456 deviceStreamManager = std::make_unique<DeviceStreamManager>(
1457 *deviceInfo, havePPDomainDecomposition(cr), runScheduleWork.simulationWork, useTiming);
1460 // If the user chose a task assignment, give them some hints
1461 // where appropriate.
1462 if (!userGpuTaskAssignment.empty())
1464 gpuTaskAssignments.logPerformanceHints(mdlog, numAvailableDevices);
1469 /* After possible communicator splitting in make_dd_communicators.
1470 * we can set up the intra/inter node communication.
1472 gmx_setup_nodecomm(fplog, cr);
1478 GMX_LOG(mdlog.warning)
1480 .appendTextFormatted(
1481 "This is simulation %d out of %d running as a composite GROMACS\n"
1482 "multi-simulation job. Setup for this simulation:\n",
1483 ms->simulationIndex_,
1484 ms->numSimulations_);
1486 GMX_LOG(mdlog.warning)
1487 .appendTextFormatted("Using %d MPI %s\n",
1490 cr->nnodes == 1 ? "thread" : "threads"
1492 cr->nnodes == 1 ? "process" : "processes"
1498 // If mdrun -pin auto honors any affinity setting that already
1499 // exists. If so, it is nice to provide feedback about whether
1500 // that existing affinity setting was from OpenMP or something
1501 // else, so we run this code both before and after we initialize
1502 // the OpenMP support.
1503 gmx_check_thread_affinity_set(mdlog, &hw_opt, hwinfo_->nthreads_hw_avail, FALSE);
1504 /* Check and update the number of OpenMP threads requested */
1505 checkAndUpdateRequestedNumOpenmpThreads(
1506 &hw_opt, *hwinfo_, cr, ms, physicalNodeComm.size_, pmeRunMode, mtop, *inputrec);
1508 gmx_omp_nthreads_init(mdlog,
1510 hwinfo_->nthreads_hw_avail,
1511 physicalNodeComm.size_,
1512 hw_opt.nthreads_omp,
1513 hw_opt.nthreads_omp_pme,
1514 !thisRankHasDuty(cr, DUTY_PP));
1516 const bool bEnableFPE = gmxShouldEnableFPExceptions();
1517 // FIXME - reconcile with gmx_feenableexcept() call from CommandLineModuleManager::run()
1520 gmx_feenableexcept();
1523 /* Now that we know the setup is consistent, check for efficiency */
1524 check_resource_division_efficiency(
1525 hwinfo_, gpuTaskAssignments.thisRankHasAnyGpuTask(), mdrunOptions.ntompOptionIsSet, cr, mdlog);
1527 /* getting number of PP/PME threads on this MPI / tMPI rank.
1528 PME: env variable should be read only on one node to make sure it is
1529 identical everywhere;
1531 const int numThreadsOnThisRank = thisRankHasDuty(cr, DUTY_PP)
1532 ? gmx_omp_nthreads_get(ModuleMultiThread::Nonbonded)
1533 : gmx_omp_nthreads_get(ModuleMultiThread::Pme);
1534 checkHardwareOversubscription(
1535 numThreadsOnThisRank, cr->nodeid, *hwinfo_->hardwareTopology, physicalNodeComm, mdlog);
1537 // Enable Peer access between GPUs where available
1538 // Only for DD, only master PP rank needs to perform setup, and only if thread MPI plus
1539 // any of the GPU communication features are active.
1540 if (DOMAINDECOMP(cr) && MASTER(cr) && thisRankHasDuty(cr, DUTY_PP) && GMX_THREAD_MPI
1541 && (runScheduleWork.simulationWork.useGpuHaloExchange
1542 || runScheduleWork.simulationWork.useGpuPmePpCommunication))
1544 setupGpuDevicePeerAccess(gpuTaskAssignments.deviceIdsAssigned(), mdlog);
1547 if (hw_opt.threadAffinity != ThreadAffinity::Off)
1549 /* Before setting affinity, check whether the affinity has changed
1550 * - which indicates that probably the OpenMP library has changed it
1551 * since we first checked).
1553 gmx_check_thread_affinity_set(mdlog, &hw_opt, hwinfo_->nthreads_hw_avail, TRUE);
1555 int numThreadsOnThisNode, intraNodeThreadOffset;
1556 analyzeThreadsOnThisNode(
1557 physicalNodeComm, numThreadsOnThisRank, &numThreadsOnThisNode, &intraNodeThreadOffset);
1559 /* Set the CPU affinity */
1560 gmx_set_thread_affinity(mdlog,
1563 *hwinfo_->hardwareTopology,
1564 numThreadsOnThisRank,
1565 numThreadsOnThisNode,
1566 intraNodeThreadOffset,
1570 if (mdrunOptions.timingOptions.resetStep > -1)
1575 "The -resetstep functionality is deprecated, and may be removed in a "
1578 std::unique_ptr<gmx_wallcycle> wcycle =
1579 wallcycle_init(fplog, mdrunOptions.timingOptions.resetStep, cr);
1583 /* Master synchronizes its value of reset_counters with all nodes
1584 * including PME only nodes */
1585 int64_t reset_counters = wcycle_get_reset_counters(wcycle.get());
1586 gmx_bcast(sizeof(reset_counters), &reset_counters, cr->mpi_comm_mysim);
1587 wcycle_set_reset_counters(wcycle.get(), reset_counters);
1590 // Membrane embedding must be initialized before we call init_forcerec()
1591 membedHolder.initializeMembed(fplog,
1598 &mdrunOptions.checkpointOptions.period);
1600 const bool thisRankHasPmeGpuTask = gpuTaskAssignments.thisRankHasPmeGpuTask();
1601 std::unique_ptr<MDAtoms> mdAtoms;
1602 std::unique_ptr<VirtualSitesHandler> vsite;
1603 std::unique_ptr<ListedForcesGpu> listedForcesGpu;
1606 if (thisRankHasDuty(cr, DUTY_PP))
1608 setupNotifier.notify(*cr);
1609 setupNotifier.notify(&atomSets);
1610 setupNotifier.notify(mtop);
1611 setupNotifier.notify(inputrec->pbcType);
1612 setupNotifier.notify(SimulationTimeStep{ inputrec->delta_t });
1613 /* Initiate forcerecord */
1614 fr = std::make_unique<t_forcerec>();
1615 fr->forceProviders = mdModules_->initForceProviders();
1616 init_forcerec(fplog,
1623 opt2fn("-table", filenames.size(), filenames.data()),
1624 opt2fn("-tablep", filenames.size(), filenames.data()),
1625 opt2fns("-tableb", filenames.size(), filenames.data()),
1627 // Dirty hack, for fixing disres and orires should be made mdmodules
1628 fr->fcdata->disres = disresdata;
1629 fr->fcdata->orires.swap(oriresData);
1631 // Save a handle to device stream manager to use elsewhere in the code
1632 // TODO: Forcerec is not a correct place to store it.
1633 fr->deviceStreamManager = deviceStreamManager.get();
1635 if (runScheduleWork.simulationWork.useGpuPmePpCommunication && !thisRankHasDuty(cr, DUTY_PME))
1638 deviceStreamManager != nullptr,
1639 "GPU device stream manager should be valid in order to use PME-PP direct "
1642 deviceStreamManager->streamIsValid(DeviceStreamType::PmePpTransfer),
1643 "GPU PP-PME stream should be valid in order to use GPU PME-PP direct "
1645 fr->pmePpCommGpu = std::make_unique<gmx::PmePpCommGpu>(
1648 deviceStreamManager->context(),
1649 deviceStreamManager->stream(DeviceStreamType::PmePpTransfer));
1652 fr->nbv = Nbnxm::init_nb_verlet(mdlog,
1657 runScheduleWork.simulationWork.useGpuNonbonded,
1658 deviceStreamManager.get(),
1662 // TODO: Move the logic below to a GPU bonded builder
1663 if (runScheduleWork.simulationWork.useGpuBonded)
1665 GMX_RELEASE_ASSERT(deviceStreamManager != nullptr,
1666 "GPU device stream manager should be valid in order to use GPU "
1667 "version of bonded forces.");
1668 listedForcesGpu = std::make_unique<ListedForcesGpu>(
1670 fr->ic->epsfac * fr->fudgeQQ,
1671 deviceStreamManager->context(),
1672 deviceStreamManager->bondedStream(havePPDomainDecomposition(cr)),
1674 fr->listedForcesGpu = listedForcesGpu.get();
1677 /* Initialize the mdAtoms structure.
1678 * mdAtoms is not filled with atom data,
1679 * as this can not be done now with domain decomposition.
1681 mdAtoms = makeMDAtoms(fplog, mtop, *inputrec, thisRankHasPmeGpuTask);
1682 if (globalState && thisRankHasPmeGpuTask)
1684 // The pinning of coordinates in the global state object works, because we only use
1685 // PME on GPU without DD or on a separate PME rank, and because the local state pointer
1686 // points to the global state object without DD.
1687 // FIXME: MD and EM separately set up the local state - this should happen in the same
1688 // function, which should also perform the pinning.
1689 changePinningPolicy(&globalState->x, pme_get_pinning_policy());
1692 /* Initialize the virtual site communication */
1693 vsite = makeVirtualSitesHandler(
1694 mtop, cr, fr->pbcType, updateGroups.updateGroupingPerMoleculeType());
1696 calc_shifts(box, fr->shift_vec);
1698 /* With periodic molecules the charge groups should be whole at start up
1699 * and the virtual sites should not be far from their proper positions.
1701 if (!inputrec->bContinuation && MASTER(cr)
1702 && !(inputrec->pbcType != PbcType::No && inputrec->bPeriodicMols))
1704 /* Make molecules whole at start of run */
1705 if (fr->pbcType != PbcType::No)
1707 do_pbc_first_mtop(fplog, inputrec->pbcType, box, &mtop, globalState->x.rvec_array());
1711 /* Correct initial vsite positions are required
1712 * for the initial distribution in the domain decomposition
1713 * and for the initial shell prediction.
1715 constructVirtualSitesGlobal(mtop, globalState->x);
1718 // Make the DD reverse topology, now that any vsites that are present are available
1719 if (DOMAINDECOMP(cr))
1721 dd_make_reverse_top(fplog, cr->dd, mtop, vsite.get(), *inputrec, domdecOptions.ddBondedChecking);
1724 if (EEL_PME(fr->ic->eeltype) || EVDW_PME(fr->ic->vdwtype))
1726 ewaldcoeff_q = fr->ic->ewaldcoeff_q;
1727 ewaldcoeff_lj = fr->ic->ewaldcoeff_lj;
1732 /* This is a PME only node */
1734 GMX_ASSERT(globalState == nullptr,
1735 "We don't need the state on a PME only rank and expect it to be unitialized");
1737 ewaldcoeff_q = calc_ewaldcoeff_q(inputrec->rcoulomb, inputrec->ewald_rtol);
1738 ewaldcoeff_lj = calc_ewaldcoeff_lj(inputrec->rvdw, inputrec->ewald_rtol_lj);
1741 gmx_pme_t* sepPmeData = nullptr;
1742 // This reference hides the fact that PME data is owned by runner on PME-only ranks and by forcerec on other ranks
1743 GMX_ASSERT(thisRankHasDuty(cr, DUTY_PP) == (fr != nullptr),
1744 "Double-checking that only PME-only ranks have no forcerec");
1745 gmx_pme_t*& pmedata = fr ? fr->pmedata : sepPmeData;
1747 // TODO should live in ewald module once its testing is improved
1749 // Later, this program could contain kernels that might be later
1750 // re-used as auto-tuning progresses, or subsequent simulations
1752 PmeGpuProgramStorage pmeGpuProgram;
1753 if (thisRankHasPmeGpuTask)
1756 (deviceStreamManager != nullptr),
1757 "GPU device stream manager should be initialized in order to use GPU for PME.");
1758 GMX_RELEASE_ASSERT((deviceInfo != nullptr),
1759 "GPU device should be initialized in order to use GPU for PME.");
1760 pmeGpuProgram = buildPmeGpuProgram(deviceStreamManager->context());
1763 /* Initiate PME if necessary,
1764 * either on all nodes or on dedicated PME nodes only. */
1765 if (EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype))
1767 if (mdAtoms && mdAtoms->mdatoms())
1769 nChargePerturbed = mdAtoms->mdatoms()->nChargePerturbed;
1770 if (EVDW_PME(inputrec->vdwtype))
1772 nTypePerturbed = mdAtoms->mdatoms()->nTypePerturbed;
1775 if (cr->npmenodes > 0)
1777 /* The PME only nodes need to know nChargePerturbed(FEP on Q) and nTypePerturbed(FEP on LJ)*/
1778 gmx_bcast(sizeof(nChargePerturbed), &nChargePerturbed, cr->mpi_comm_mysim);
1779 gmx_bcast(sizeof(nTypePerturbed), &nTypePerturbed, cr->mpi_comm_mysim);
1782 if (thisRankHasDuty(cr, DUTY_PME))
1786 // TODO: This should be in the builder.
1787 GMX_RELEASE_ASSERT(!runScheduleWork.simulationWork.useGpuPme
1788 || (deviceStreamManager != nullptr),
1789 "Device stream manager should be valid in order to use GPU "
1792 !runScheduleWork.simulationWork.useGpuPme
1793 || deviceStreamManager->streamIsValid(DeviceStreamType::Pme),
1794 "GPU PME stream should be valid in order to use GPU version of PME.");
1796 const DeviceContext* deviceContext = runScheduleWork.simulationWork.useGpuPme
1797 ? &deviceStreamManager->context()
1799 const DeviceStream* pmeStream =
1800 runScheduleWork.simulationWork.useGpuPme
1801 ? &deviceStreamManager->stream(DeviceStreamType::Pme)
1804 pmedata = gmx_pme_init(cr,
1805 getNumPmeDomains(cr->dd),
1807 nChargePerturbed != 0,
1808 nTypePerturbed != 0,
1809 mdrunOptions.reproducible,
1812 gmx_omp_nthreads_get(ModuleMultiThread::Pme),
1817 pmeGpuProgram.get(),
1820 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1825 if (EI_DYNAMICS(inputrec->eI))
1827 /* Turn on signal handling on all nodes */
1829 * (A user signal from the PME nodes (if any)
1830 * is communicated to the PP nodes.
1832 signal_handler_install();
1835 pull_t* pull_work = nullptr;
1836 if (thisRankHasDuty(cr, DUTY_PP))
1838 /* Assumes uniform use of the number of OpenMP threads */
1839 walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(ModuleMultiThread::Default));
1841 if (inputrec->bPull)
1843 /* Initialize pull code */
1844 pull_work = init_pull(fplog,
1845 inputrec->pull.get(),
1850 inputrec->fepvals->init_lambda);
1851 if (inputrec->pull->bXOutAverage || inputrec->pull->bFOutAverage)
1853 initPullHistory(pull_work, &observablesHistory);
1855 if (EI_DYNAMICS(inputrec->eI) && MASTER(cr))
1857 init_pull_output_files(pull_work, filenames.size(), filenames.data(), oenv, startingBehavior);
1861 std::unique_ptr<EnforcedRotation> enforcedRotation;
1864 /* Initialize enforced rotation code */
1865 enforcedRotation = init_rot(fplog,
1878 t_swap* swap = nullptr;
1879 if (inputrec->eSwapCoords != SwapType::No)
1881 /* Initialize ion swapping code */
1882 swap = init_swapcoords(fplog,
1884 opt2fn_master("-swap", filenames.size(), filenames.data(), cr),
1887 &observablesHistory,
1895 /* Let makeConstraints know whether we have essential dynamics constraints. */
1896 auto constr = makeConstraints(mtop,
1899 doEssentialDynamics,
1902 updateGroups.useUpdateGroups(),
1908 /* Energy terms and groups */
1909 gmx_enerdata_t enerd(mtop.groups.groups[SimulationAtomGroupType::EnergyOutput].size(),
1910 inputrec->fepvals->n_lambda);
1912 // cos acceleration is only supported by md, but older tpr
1913 // files might still combine it with other integrators
1914 GMX_RELEASE_ASSERT(inputrec->cos_accel == 0.0 || inputrec->eI == IntegrationAlgorithm::MD,
1915 "cos_acceleration is only supported by integrator=md");
1917 /* Kinetic energy data */
1918 gmx_ekindata_t ekind(inputrec->opts.ngtc,
1919 inputrec->cos_accel,
1920 gmx_omp_nthreads_get(ModuleMultiThread::Update));
1922 /* Set up interactive MD (IMD) */
1923 auto imdSession = makeImdSession(inputrec.get(),
1930 MASTER(cr) ? globalState->x : gmx::ArrayRef<gmx::RVec>(),
1934 mdrunOptions.imdOptions,
1937 if (DOMAINDECOMP(cr))
1939 GMX_RELEASE_ASSERT(fr, "fr was NULL while cr->duty was DUTY_PP");
1940 /* This call is not included in init_domain_decomposition
1941 * because fr->cginfo_mb is set later.
1943 makeBondedLinks(cr->dd, mtop, fr->cginfo_mb);
1946 if (runScheduleWork.simulationWork.useGpuBufferOps)
1948 fr->gpuForceReduction[gmx::AtomLocality::Local] = std::make_unique<gmx::GpuForceReduction>(
1949 deviceStreamManager->context(),
1950 deviceStreamManager->stream(gmx::DeviceStreamType::NonBondedLocal),
1952 fr->gpuForceReduction[gmx::AtomLocality::NonLocal] = std::make_unique<gmx::GpuForceReduction>(
1953 deviceStreamManager->context(),
1954 deviceStreamManager->stream(gmx::DeviceStreamType::NonBondedNonLocal),
1958 std::unique_ptr<gmx::StatePropagatorDataGpu> stateGpu;
1959 if (gpusWereDetected
1960 && ((runScheduleWork.simulationWork.useGpuPme && thisRankHasDuty(cr, DUTY_PME))
1961 || runScheduleWork.simulationWork.useGpuBufferOps))
1963 GpuApiCallBehavior transferKind =
1964 (inputrec->eI == IntegrationAlgorithm::MD && !doRerun && !useModularSimulator)
1965 ? GpuApiCallBehavior::Async
1966 : GpuApiCallBehavior::Sync;
1967 GMX_RELEASE_ASSERT(deviceStreamManager != nullptr,
1968 "GPU device stream manager should be initialized to use GPU.");
1969 stateGpu = std::make_unique<gmx::StatePropagatorDataGpu>(
1970 *deviceStreamManager, transferKind, pme_gpu_get_block_size(fr->pmedata), wcycle.get());
1971 fr->stateGpu = stateGpu.get();
1974 GMX_ASSERT(stopHandlerBuilder_, "Runner must provide StopHandlerBuilder to simulator.");
1975 SimulatorBuilder simulatorBuilder;
1977 simulatorBuilder.add(SimulatorStateData(globalState.get(), &observablesHistory, &enerd, &ekind));
1978 simulatorBuilder.add(std::move(membedHolder));
1979 simulatorBuilder.add(std::move(stopHandlerBuilder_));
1980 simulatorBuilder.add(SimulatorConfig(mdrunOptions, startingBehavior, &runScheduleWork));
1983 simulatorBuilder.add(SimulatorEnv(fplog, cr, ms, mdlog, oenv));
1984 simulatorBuilder.add(Profiling(&nrnb, walltime_accounting, wcycle.get()));
1985 simulatorBuilder.add(ConstraintsParam(
1986 constr.get(), enforcedRotation ? enforcedRotation->getLegacyEnfrot() : nullptr, vsite.get()));
1987 // TODO: Separate `fr` to a separate add, and make the `build` handle the coupling sensibly.
1988 simulatorBuilder.add(LegacyInput(
1989 static_cast<int>(filenames.size()), filenames.data(), inputrec.get(), fr.get()));
1990 simulatorBuilder.add(ReplicaExchangeParameters(replExParams));
1991 simulatorBuilder.add(InteractiveMD(imdSession.get()));
1992 simulatorBuilder.add(SimulatorModules(mdModules_->outputProvider(), mdModules_->notifiers()));
1993 simulatorBuilder.add(CenterOfMassPulling(pull_work));
1994 // Todo move to an MDModule
1995 simulatorBuilder.add(IonSwapping(swap));
1996 simulatorBuilder.add(TopologyData(mtop, mdAtoms.get()));
1997 simulatorBuilder.add(BoxDeformationHandle(deform.get()));
1998 simulatorBuilder.add(std::move(modularSimulatorCheckpointData));
2000 // build and run simulator object based on user-input
2001 auto simulator = simulatorBuilder.build(useModularSimulator);
2004 if (fr->pmePpCommGpu)
2006 // destroy object since it is no longer required. (This needs to be done while the GPU context still exists.)
2007 fr->pmePpCommGpu.reset();
2010 if (inputrec->bPull)
2012 finish_pull(pull_work);
2014 finish_swapcoords(swap);
2018 GMX_RELEASE_ASSERT(pmedata, "pmedata was NULL while cr->duty was not DUTY_PP");
2020 walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(ModuleMultiThread::Pme));
2021 gmx_pmeonly(pmedata,
2025 walltime_accounting,
2028 runScheduleWork.simulationWork.useGpuPmePpCommunication,
2029 deviceStreamManager.get());
2032 wallcycle_stop(wcycle.get(), WallCycleCounter::Run);
2034 /* Finish up, write some stuff
2035 * if rerunMD, don't write last frame again
2043 walltime_accounting,
2044 fr ? fr->nbv.get() : nullptr,
2046 EI_DYNAMICS(inputrec->eI) && !isMultiSim(ms));
2049 deviceStreamManager.reset(nullptr);
2053 gmx_pme_destroy(pmedata);
2057 // FIXME: this is only here to manually unpin mdAtoms->chargeA_ and state->x,
2058 // before we destroy the GPU context(s)
2059 // Pinned buffers are associated with contexts in CUDA.
2060 // As soon as we destroy GPU contexts after mdrunner() exits, these lines should go.
2061 mdAtoms.reset(nullptr);
2062 globalState.reset(nullptr);
2063 mdModules_.reset(nullptr); // destruct force providers here as they might also use the GPU
2064 listedForcesGpu.reset(nullptr);
2065 fr.reset(nullptr); // destruct forcerec before gpu
2066 // TODO convert to C++ so we can get rid of these frees
2069 if (!hwinfo_->deviceInfoList.empty())
2071 /* stop the GPU profiler (only CUDA) */
2075 /* With tMPI we need to wait for all ranks to finish deallocation before
2076 * destroying the CUDA context as some tMPI ranks may be sharing
2079 * This is not a concern in OpenCL where we use one context per rank.
2081 * Note: it is safe to not call the barrier on the ranks which do not use GPU,
2082 * but it is easier and more futureproof to call it on the whole node.
2084 * Note that this function needs to be called even if GPUs are not used
2085 * in this run because the PME ranks have no knowledge of whether GPUs
2086 * are used or not, but all ranks need to enter the barrier below.
2087 * \todo Remove this physical node barrier after making sure
2088 * that it's not needed anymore (with a shared GPU run).
2092 physicalNodeComm.barrier();
2095 if (!devFlags.usingCudaAwareMpi)
2097 // Don't reset GPU in case of CUDA-AWARE MPI
2098 // UCX creates CUDA buffers which are cleaned-up as part of MPI_Finalize()
2099 // resetting the device before MPI_Finalize() results in crashes inside UCX
2100 releaseDevice(deviceInfo);
2103 /* Does what it says */
2104 print_date_and_time(fplog, cr->nodeid, "Finished mdrun", gmx_gettime());
2105 walltime_accounting_destroy(walltime_accounting);
2107 // Ensure log file content is written
2110 gmx_fio_flush(logFileHandle);
2113 /* Reset FPEs (important for unit tests) by disabling them. Assumes no
2114 * exceptions were enabled before function was called. */
2117 gmx_fedisableexcept();
2120 auto rc = static_cast<int>(gmx_get_stop_condition());
2123 /* we need to join all threads. The sub-threads join when they
2124 exit this function, but the master thread needs to be told to
2134 Mdrunner::~Mdrunner()
2136 // Clean up of the Manager.
2137 // This will end up getting called on every thread-MPI rank, which is unnecessary,
2138 // but okay as long as threads synchronize some time before adding or accessing
2139 // a new set of restraints.
2140 if (restraintManager_)
2142 restraintManager_->clear();
2143 GMX_ASSERT(restraintManager_->countRestraints() == 0,
2144 "restraints added during runner life time should be cleared at runner "
2149 void Mdrunner::addPotential(std::shared_ptr<gmx::IRestraintPotential> puller, const std::string& name)
2151 GMX_ASSERT(restraintManager_, "Mdrunner must have a restraint manager.");
2152 // Not sure if this should be logged through the md logger or something else,
2153 // but it is helpful to have some sort of INFO level message sent somewhere.
2154 // std::cout << "Registering restraint named " << name << std::endl;
2156 // When multiple restraints are used, it may be wasteful to register them separately.
2157 // Maybe instead register an entire Restraint Manager as a force provider.
2158 restraintManager_->addToSpec(std::move(puller), name);
2161 Mdrunner::Mdrunner(std::unique_ptr<MDModules> mdModules) : mdModules_(std::move(mdModules)) {}
2163 Mdrunner::Mdrunner(Mdrunner&&) noexcept = default;
2165 //NOLINTNEXTLINE(performance-noexcept-move-constructor) working around GCC bug 58265 in CentOS 7
2166 Mdrunner& Mdrunner::operator=(Mdrunner&& /*handle*/) noexcept(BUGFREE_NOEXCEPT_STRING) = default;
2168 class Mdrunner::BuilderImplementation
2171 BuilderImplementation() = delete;
2172 BuilderImplementation(std::unique_ptr<MDModules> mdModules, compat::not_null<SimulationContext*> context);
2173 ~BuilderImplementation();
2175 BuilderImplementation& setExtraMdrunOptions(const MdrunOptions& options,
2176 real forceWarningThreshold,
2177 StartingBehavior startingBehavior);
2179 void addHardwareDetectionResult(const gmx_hw_info_t* hwinfo);
2181 void addDomdec(const DomdecOptions& options);
2183 void addInput(SimulationInputHandle inputHolder);
2185 void addVerletList(int nstlist);
2187 void addReplicaExchange(const ReplicaExchangeParameters& params);
2189 void addNonBonded(const char* nbpu_opt);
2191 void addPME(const char* pme_opt_, const char* pme_fft_opt_);
2193 void addBondedTaskAssignment(const char* bonded_opt);
2195 void addUpdateTaskAssignment(const char* update_opt);
2197 void addHardwareOptions(const gmx_hw_opt_t& hardwareOptions);
2199 void addFilenames(ArrayRef<const t_filenm> filenames);
2201 void addOutputEnvironment(gmx_output_env_t* outputEnvironment);
2203 void addLogFile(t_fileio* logFileHandle);
2205 void addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder);
2210 // Default parameters copied from runner.h
2211 // \todo Clarify source(s) of default parameters.
2213 const char* nbpu_opt_ = nullptr;
2214 const char* pme_opt_ = nullptr;
2215 const char* pme_fft_opt_ = nullptr;
2216 const char* bonded_opt_ = nullptr;
2217 const char* update_opt_ = nullptr;
2219 MdrunOptions mdrunOptions_;
2221 DomdecOptions domdecOptions_;
2223 ReplicaExchangeParameters replicaExchangeParameters_;
2225 //! Command-line override for the duration of a neighbor list with the Verlet scheme.
2228 //! World communicator, used for hardware detection and task assignment
2229 MPI_Comm libraryWorldCommunicator_ = MPI_COMM_NULL;
2231 //! Multisim communicator handle.
2232 gmx_multisim_t* multiSimulation_;
2234 //! mdrun communicator
2235 MPI_Comm simulationCommunicator_ = MPI_COMM_NULL;
2237 //! Print a warning if any force is larger than this (in kJ/mol nm).
2238 real forceWarningThreshold_ = -1;
2240 //! Whether the simulation will start afresh, or restart with/without appending.
2241 StartingBehavior startingBehavior_ = StartingBehavior::NewSimulation;
2243 //! The modules that comprise the functionality of mdrun.
2244 std::unique_ptr<MDModules> mdModules_;
2246 //! Detected hardware.
2247 const gmx_hw_info_t* hwinfo_ = nullptr;
2249 //! \brief Parallelism information.
2250 gmx_hw_opt_t hardwareOptions_;
2252 //! filename options for simulation.
2253 ArrayRef<const t_filenm> filenames_;
2255 /*! \brief Handle to output environment.
2257 * \todo gmx_output_env_t needs lifetime management.
2259 gmx_output_env_t* outputEnvironment_ = nullptr;
2261 /*! \brief Non-owning handle to MD log file.
2263 * \todo Context should own output facilities for client.
2264 * \todo Improve log file handle management.
2266 * Code managing the FILE* relies on the ability to set it to
2267 * nullptr to check whether the filehandle is valid.
2269 t_fileio* logFileHandle_ = nullptr;
2272 * \brief Builder for simulation stop signal handler.
2274 std::unique_ptr<StopHandlerBuilder> stopHandlerBuilder_ = nullptr;
2277 * \brief Sources for initial simulation state.
2279 * See issue #3652 for near-term refinements to the SimulationInput interface.
2281 * See issue #3379 for broader discussion on API aspects of simulation inputs and outputs.
2283 SimulationInputHandle inputHolder_;
2286 Mdrunner::BuilderImplementation::BuilderImplementation(std::unique_ptr<MDModules> mdModules,
2287 compat::not_null<SimulationContext*> context) :
2288 mdModules_(std::move(mdModules))
2290 libraryWorldCommunicator_ = context->libraryWorldCommunicator_;
2291 simulationCommunicator_ = context->simulationCommunicator_;
2292 multiSimulation_ = context->multiSimulation_.get();
2295 Mdrunner::BuilderImplementation::~BuilderImplementation() = default;
2297 Mdrunner::BuilderImplementation&
2298 Mdrunner::BuilderImplementation::setExtraMdrunOptions(const MdrunOptions& options,
2299 const real forceWarningThreshold,
2300 const StartingBehavior startingBehavior)
2302 mdrunOptions_ = options;
2303 forceWarningThreshold_ = forceWarningThreshold;
2304 startingBehavior_ = startingBehavior;
2308 void Mdrunner::BuilderImplementation::addDomdec(const DomdecOptions& options)
2310 domdecOptions_ = options;
2313 void Mdrunner::BuilderImplementation::addVerletList(int nstlist)
2318 void Mdrunner::BuilderImplementation::addReplicaExchange(const ReplicaExchangeParameters& params)
2320 replicaExchangeParameters_ = params;
2323 Mdrunner Mdrunner::BuilderImplementation::build()
2325 auto newRunner = Mdrunner(std::move(mdModules_));
2327 newRunner.mdrunOptions = mdrunOptions_;
2328 newRunner.pforce = forceWarningThreshold_;
2329 newRunner.startingBehavior = startingBehavior_;
2330 newRunner.domdecOptions = domdecOptions_;
2332 // \todo determine an invariant to check or confirm that all gmx_hw_opt_t objects are valid
2333 newRunner.hw_opt = hardwareOptions_;
2335 // No invariant to check. This parameter exists to optionally override other behavior.
2336 newRunner.nstlist_cmdline = nstlist_;
2338 newRunner.replExParams = replicaExchangeParameters_;
2340 newRunner.filenames = filenames_;
2342 newRunner.libraryWorldCommunicator = libraryWorldCommunicator_;
2344 newRunner.simulationCommunicator = simulationCommunicator_;
2346 // nullptr is a valid value for the multisim handle
2347 newRunner.ms = multiSimulation_;
2351 newRunner.hwinfo_ = hwinfo_;
2355 GMX_THROW(gmx::APIError(
2356 "MdrunnerBuilder::addHardwareDetectionResult() is required before build()"));
2361 newRunner.inputHolder_ = std::move(inputHolder_);
2365 GMX_THROW(gmx::APIError("MdrunnerBuilder::addInput() is required before build()."));
2368 // \todo Clarify ownership and lifetime management for gmx_output_env_t
2369 // \todo Update sanity checking when output environment has clearly specified invariants.
2370 // Initialization and default values for oenv are not well specified in the current version.
2371 if (outputEnvironment_)
2373 newRunner.oenv = outputEnvironment_;
2377 GMX_THROW(gmx::APIError(
2378 "MdrunnerBuilder::addOutputEnvironment() is required before build()"));
2381 newRunner.logFileHandle = logFileHandle_;
2385 newRunner.nbpu_opt = nbpu_opt_;
2389 GMX_THROW(gmx::APIError("MdrunnerBuilder::addNonBonded() is required before build()"));
2392 if (pme_opt_ && pme_fft_opt_)
2394 newRunner.pme_opt = pme_opt_;
2395 newRunner.pme_fft_opt = pme_fft_opt_;
2399 GMX_THROW(gmx::APIError("MdrunnerBuilder::addElectrostatics() is required before build()"));
2404 newRunner.bonded_opt = bonded_opt_;
2408 GMX_THROW(gmx::APIError(
2409 "MdrunnerBuilder::addBondedTaskAssignment() is required before build()"));
2414 newRunner.update_opt = update_opt_;
2418 GMX_THROW(gmx::APIError(
2419 "MdrunnerBuilder::addUpdateTaskAssignment() is required before build() "));
2423 newRunner.restraintManager_ = std::make_unique<gmx::RestraintManager>();
2425 if (stopHandlerBuilder_)
2427 newRunner.stopHandlerBuilder_ = std::move(stopHandlerBuilder_);
2431 newRunner.stopHandlerBuilder_ = std::make_unique<StopHandlerBuilder>();
2437 void Mdrunner::BuilderImplementation::addHardwareDetectionResult(const gmx_hw_info_t* hwinfo)
2442 void Mdrunner::BuilderImplementation::addNonBonded(const char* nbpu_opt)
2444 nbpu_opt_ = nbpu_opt;
2447 void Mdrunner::BuilderImplementation::addPME(const char* pme_opt, const char* pme_fft_opt)
2450 pme_fft_opt_ = pme_fft_opt;
2453 void Mdrunner::BuilderImplementation::addBondedTaskAssignment(const char* bonded_opt)
2455 bonded_opt_ = bonded_opt;
2458 void Mdrunner::BuilderImplementation::addUpdateTaskAssignment(const char* update_opt)
2460 update_opt_ = update_opt;
2463 void Mdrunner::BuilderImplementation::addHardwareOptions(const gmx_hw_opt_t& hardwareOptions)
2465 hardwareOptions_ = hardwareOptions;
2468 void Mdrunner::BuilderImplementation::addFilenames(ArrayRef<const t_filenm> filenames)
2470 filenames_ = filenames;
2473 void Mdrunner::BuilderImplementation::addOutputEnvironment(gmx_output_env_t* outputEnvironment)
2475 outputEnvironment_ = outputEnvironment;
2478 void Mdrunner::BuilderImplementation::addLogFile(t_fileio* logFileHandle)
2480 logFileHandle_ = logFileHandle;
2483 void Mdrunner::BuilderImplementation::addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder)
2485 stopHandlerBuilder_ = std::move(builder);
2488 void Mdrunner::BuilderImplementation::addInput(SimulationInputHandle inputHolder)
2490 inputHolder_ = std::move(inputHolder);
2493 MdrunnerBuilder::MdrunnerBuilder(std::unique_ptr<MDModules> mdModules,
2494 compat::not_null<SimulationContext*> context) :
2495 impl_{ std::make_unique<Mdrunner::BuilderImplementation>(std::move(mdModules), context) }
2499 MdrunnerBuilder::~MdrunnerBuilder() = default;
2501 MdrunnerBuilder& MdrunnerBuilder::addHardwareDetectionResult(const gmx_hw_info_t* hwinfo)
2503 impl_->addHardwareDetectionResult(hwinfo);
2507 MdrunnerBuilder& MdrunnerBuilder::addSimulationMethod(const MdrunOptions& options,
2508 real forceWarningThreshold,
2509 const StartingBehavior startingBehavior)
2511 impl_->setExtraMdrunOptions(options, forceWarningThreshold, startingBehavior);
2515 MdrunnerBuilder& MdrunnerBuilder::addDomainDecomposition(const DomdecOptions& options)
2517 impl_->addDomdec(options);
2521 MdrunnerBuilder& MdrunnerBuilder::addNeighborList(int nstlist)
2523 impl_->addVerletList(nstlist);
2527 MdrunnerBuilder& MdrunnerBuilder::addReplicaExchange(const ReplicaExchangeParameters& params)
2529 impl_->addReplicaExchange(params);
2533 MdrunnerBuilder& MdrunnerBuilder::addNonBonded(const char* nbpu_opt)
2535 impl_->addNonBonded(nbpu_opt);
2539 MdrunnerBuilder& MdrunnerBuilder::addElectrostatics(const char* pme_opt, const char* pme_fft_opt)
2541 // The builder method may become more general in the future, but in this version,
2542 // parameters for PME electrostatics are both required and the only parameters
2544 if (pme_opt && pme_fft_opt)
2546 impl_->addPME(pme_opt, pme_fft_opt);
2551 gmx::InvalidInputError("addElectrostatics() arguments must be non-null pointers."));
2556 MdrunnerBuilder& MdrunnerBuilder::addBondedTaskAssignment(const char* bonded_opt)
2558 impl_->addBondedTaskAssignment(bonded_opt);
2562 MdrunnerBuilder& MdrunnerBuilder::addUpdateTaskAssignment(const char* update_opt)
2564 impl_->addUpdateTaskAssignment(update_opt);
2568 Mdrunner MdrunnerBuilder::build()
2570 return impl_->build();
2573 MdrunnerBuilder& MdrunnerBuilder::addHardwareOptions(const gmx_hw_opt_t& hardwareOptions)
2575 impl_->addHardwareOptions(hardwareOptions);
2579 MdrunnerBuilder& MdrunnerBuilder::addFilenames(ArrayRef<const t_filenm> filenames)
2581 impl_->addFilenames(filenames);
2585 MdrunnerBuilder& MdrunnerBuilder::addOutputEnvironment(gmx_output_env_t* outputEnvironment)
2587 impl_->addOutputEnvironment(outputEnvironment);
2591 MdrunnerBuilder& MdrunnerBuilder::addLogFile(t_fileio* logFileHandle)
2593 impl_->addLogFile(logFileHandle);
2597 MdrunnerBuilder& MdrunnerBuilder::addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder)
2599 impl_->addStopHandlerBuilder(std::move(builder));
2603 MdrunnerBuilder& MdrunnerBuilder::addInput(SimulationInputHandle input)
2605 impl_->addInput(std::move(input));
2609 MdrunnerBuilder::MdrunnerBuilder(MdrunnerBuilder&&) noexcept = default;
2611 MdrunnerBuilder& MdrunnerBuilder::operator=(MdrunnerBuilder&&) noexcept = default;