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 // Some builds of GCC 5 give false positive warnings that these
206 // getenv results are ignored when clearly they are used.
207 #pragma GCC diagnostic push
208 #pragma GCC diagnostic ignored "-Wunused-result"
210 devFlags.enableGpuBufferOps =
211 GMX_GPU_CUDA && useGpuForNonbonded && (getenv("GMX_USE_GPU_BUFFER_OPS") != nullptr);
212 devFlags.enableGpuHaloExchange = GMX_GPU_CUDA && getenv("GMX_GPU_DD_COMMS") != nullptr;
213 devFlags.forceGpuUpdateDefault = (getenv("GMX_FORCE_UPDATE_DEFAULT_GPU") != nullptr) || GMX_FAHCORE;
214 devFlags.enableGpuPmePPComm = GMX_GPU_CUDA && getenv("GMX_GPU_PME_PP_COMMS") != nullptr;
216 #pragma GCC diagnostic pop
218 // Direct GPU comm path is being used with CUDA_AWARE_MPI
219 // make sure underlying MPI implementation is CUDA-aware
220 if (!GMX_THREAD_MPI && (devFlags.enableGpuPmePPComm || devFlags.enableGpuHaloExchange))
222 const bool haveDetectedCudaAwareMpi =
223 (checkMpiCudaAwareSupport() == CudaAwareMpiStatus::Supported);
224 const bool forceCudaAwareMpi = (getenv("GMX_FORCE_CUDA_AWARE_MPI") != nullptr);
226 if (!haveDetectedCudaAwareMpi && forceCudaAwareMpi)
228 // CUDA-aware support not detected in MPI library but, user has forced it's use
229 GMX_LOG(mdlog.warning)
231 .appendTextFormatted(
232 "This run has forced use of 'CUDA-aware MPI'. "
233 "But, GROMACS cannot determine if underlying MPI "
234 "is CUDA-aware. GROMACS recommends use of latest openMPI version "
235 "for CUDA-aware support. "
236 "If you observe failures at runtime, try unsetting "
237 "GMX_FORCE_CUDA_AWARE_MPI environment variable.");
240 if (haveDetectedCudaAwareMpi || forceCudaAwareMpi)
242 devFlags.usingCudaAwareMpi = true;
243 GMX_LOG(mdlog.warning)
245 .appendTextFormatted(
246 "Using CUDA-aware MPI for 'GPU halo exchange' or 'GPU PME-PP "
247 "communications' feature.");
251 if (devFlags.enableGpuHaloExchange)
253 GMX_LOG(mdlog.warning)
255 .appendTextFormatted(
256 "GMX_GPU_DD_COMMS environment variable detected, but the 'GPU "
257 "halo exchange' feature will not be enabled as GROMACS couldn't "
258 "detect CUDA_aware support in underlying MPI implementation.");
259 devFlags.enableGpuHaloExchange = false;
261 if (devFlags.enableGpuPmePPComm)
263 GMX_LOG(mdlog.warning)
266 "GMX_GPU_PME_PP_COMMS environment variable detected, but the "
267 "'GPU PME-PP communications' feature will not be enabled as "
269 "detect CUDA_aware support in underlying MPI implementation.");
270 devFlags.enableGpuPmePPComm = false;
273 GMX_LOG(mdlog.warning)
275 .appendTextFormatted(
276 "GROMACS recommends use of latest OpenMPI version for CUDA-aware "
278 "If you are certain about CUDA-aware support in your MPI library, "
279 "you can force it's use by setting environment variable "
280 " GMX_FORCE_CUDA_AWARE_MPI.");
284 if (devFlags.enableGpuBufferOps)
286 GMX_LOG(mdlog.warning)
288 .appendTextFormatted(
289 "This run uses the 'GPU buffer ops' feature, enabled by the "
290 "GMX_USE_GPU_BUFFER_OPS environment variable.");
293 if (devFlags.forceGpuUpdateDefault)
295 GMX_LOG(mdlog.warning)
297 .appendTextFormatted(
298 "This run will default to '-update gpu' as requested by the "
299 "GMX_FORCE_UPDATE_DEFAULT_GPU environment variable. GPU update with domain "
300 "decomposition lacks substantial testing and should be used with caution.");
303 if (devFlags.enableGpuHaloExchange)
305 if (useGpuForNonbonded)
307 if (!devFlags.enableGpuBufferOps)
309 GMX_LOG(mdlog.warning)
311 .appendTextFormatted(
312 "Enabling GPU buffer operations required by GMX_GPU_DD_COMMS "
313 "(equivalent with GMX_USE_GPU_BUFFER_OPS=1).");
314 devFlags.enableGpuBufferOps = true;
316 GMX_LOG(mdlog.warning)
318 .appendTextFormatted(
319 "This run has requested the 'GPU halo exchange' feature, enabled by "
321 "GMX_GPU_DD_COMMS environment variable.");
325 GMX_LOG(mdlog.warning)
327 .appendTextFormatted(
328 "GMX_GPU_DD_COMMS environment variable detected, but the 'GPU "
329 "halo exchange' feature will not be enabled as nonbonded interactions "
330 "are not offloaded.");
331 devFlags.enableGpuHaloExchange = false;
335 if (devFlags.enableGpuPmePPComm)
337 if (pmeRunMode == PmeRunMode::GPU)
339 if (!devFlags.enableGpuBufferOps)
341 GMX_LOG(mdlog.warning)
343 .appendTextFormatted(
344 "Enabling GPU buffer operations required by GMX_GPU_PME_PP_COMMS "
345 "(equivalent with GMX_USE_GPU_BUFFER_OPS=1).");
346 devFlags.enableGpuBufferOps = true;
348 GMX_LOG(mdlog.warning)
350 .appendTextFormatted(
351 "This run uses the 'GPU PME-PP communications' feature, enabled "
352 "by the GMX_GPU_PME_PP_COMMS environment variable.");
356 std::string clarification;
357 if (pmeRunMode == PmeRunMode::Mixed)
360 "PME FFT and gather are not offloaded to the GPU (PME is running in mixed "
365 clarification = "PME is not offloaded to the GPU.";
367 GMX_LOG(mdlog.warning)
370 "GMX_GPU_PME_PP_COMMS environment variable detected, but the "
371 "'GPU PME-PP communications' feature was not enabled as "
373 devFlags.enableGpuPmePPComm = false;
380 /*! \brief Barrier for safe simultaneous thread access to mdrunner data
382 * Used to ensure that the master thread does not modify mdrunner during copy
383 * on the spawned threads. */
384 static void threadMpiMdrunnerAccessBarrier()
387 MPI_Barrier(MPI_COMM_WORLD);
391 Mdrunner Mdrunner::cloneOnSpawnedThread() const
393 auto newRunner = Mdrunner(std::make_unique<MDModules>());
395 // All runners in the same process share a restraint manager resource because it is
396 // part of the interface to the client code, which is associated only with the
397 // original thread. Handles to the same resources can be obtained by copy.
399 newRunner.restraintManager_ = std::make_unique<RestraintManager>(*restraintManager_);
402 // Copy members of master runner.
403 // \todo Replace with builder when Simulation context and/or runner phases are better defined.
404 // Ref https://gitlab.com/gromacs/gromacs/-/issues/2587 and https://gitlab.com/gromacs/gromacs/-/issues/2375
405 newRunner.hw_opt = hw_opt;
406 newRunner.filenames = filenames;
408 newRunner.hwinfo_ = hwinfo_;
409 newRunner.oenv = oenv;
410 newRunner.mdrunOptions = mdrunOptions;
411 newRunner.domdecOptions = domdecOptions;
412 newRunner.nbpu_opt = nbpu_opt;
413 newRunner.pme_opt = pme_opt;
414 newRunner.pme_fft_opt = pme_fft_opt;
415 newRunner.bonded_opt = bonded_opt;
416 newRunner.update_opt = update_opt;
417 newRunner.nstlist_cmdline = nstlist_cmdline;
418 newRunner.replExParams = replExParams;
419 newRunner.pforce = pforce;
420 // Give the spawned thread the newly created valid communicator
421 // for the simulation.
422 newRunner.libraryWorldCommunicator = MPI_COMM_WORLD;
423 newRunner.simulationCommunicator = MPI_COMM_WORLD;
425 newRunner.startingBehavior = startingBehavior;
426 newRunner.stopHandlerBuilder_ = std::make_unique<StopHandlerBuilder>(*stopHandlerBuilder_);
427 newRunner.inputHolder_ = inputHolder_;
429 threadMpiMdrunnerAccessBarrier();
434 /*! \brief The callback used for running on spawned threads.
436 * Obtains the pointer to the master mdrunner object from the one
437 * argument permitted to the thread-launch API call, copies it to make
438 * a new runner for this thread, reinitializes necessary data, and
439 * proceeds to the simulation. */
440 static void mdrunner_start_fn(const void* arg)
444 const auto* masterMdrunner = reinterpret_cast<const gmx::Mdrunner*>(arg);
445 /* copy the arg list to make sure that it's thread-local. This
446 doesn't copy pointed-to items, of course; fnm, cr and fplog
447 are reset in the call below, all others should be const. */
448 gmx::Mdrunner mdrunner = masterMdrunner->cloneOnSpawnedThread();
451 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
455 void Mdrunner::spawnThreads(int numThreadsToLaunch)
458 /* now spawn new threads that start mdrunner_start_fn(), while
459 the main thread returns. Thread affinity is handled later. */
460 if (tMPI_Init_fn(TRUE, numThreadsToLaunch, TMPI_AFFINITY_NONE, mdrunner_start_fn, static_cast<const void*>(this))
463 GMX_THROW(gmx::InternalError("Failed to spawn thread-MPI threads"));
466 // Give the master thread the newly created valid communicator for
468 libraryWorldCommunicator = MPI_COMM_WORLD;
469 simulationCommunicator = MPI_COMM_WORLD;
470 threadMpiMdrunnerAccessBarrier();
472 GMX_UNUSED_VALUE(numThreadsToLaunch);
473 GMX_UNUSED_VALUE(mdrunner_start_fn);
479 /*! \brief Initialize variables for Verlet scheme simulation */
480 static void prepare_verlet_scheme(FILE* fplog,
484 const gmx_mtop_t& mtop,
486 bool makeGpuPairList,
487 const gmx::CpuInfo& cpuinfo)
489 // We checked the cut-offs in grompp, but double-check here.
490 // We have PME+LJcutoff kernels for rcoulomb>rvdw.
491 if (EEL_PME_EWALD(ir->coulombtype) && ir->vdwtype == VanDerWaalsType::Cut)
493 GMX_RELEASE_ASSERT(ir->rcoulomb >= ir->rvdw,
494 "With Verlet lists and PME we should have rcoulomb>=rvdw");
498 GMX_RELEASE_ASSERT(ir->rcoulomb == ir->rvdw,
499 "With Verlet lists and no PME rcoulomb and rvdw should be identical");
501 /* For NVE simulations, we will retain the initial list buffer */
502 if (EI_DYNAMICS(ir->eI) && ir->verletbuf_tol > 0
503 && !(EI_MD(ir->eI) && ir->etc == TemperatureCoupling::No))
505 /* Update the Verlet buffer size for the current run setup */
507 /* Here we assume SIMD-enabled kernels are being used. But as currently
508 * calc_verlet_buffer_size gives the same results for 4x8 and 4x4
509 * and 4x2 gives a larger buffer than 4x4, this is ok.
511 ListSetupType listType =
512 (makeGpuPairList ? ListSetupType::Gpu : ListSetupType::CpuSimdWhenSupported);
513 VerletbufListSetup listSetup = verletbufGetSafeListSetup(listType);
515 const real rlist_new =
516 calcVerletBufferSize(mtop, det(box), *ir, ir->nstlist, ir->nstlist - 1, -1, listSetup);
518 if (rlist_new != ir->rlist)
520 if (fplog != nullptr)
523 "\nChanging rlist from %g to %g for non-bonded %dx%d atom kernels\n\n",
526 listSetup.cluster_size_i,
527 listSetup.cluster_size_j);
529 ir->rlist = rlist_new;
533 if (nstlist_cmdline > 0 && (!EI_DYNAMICS(ir->eI) || ir->verletbuf_tol <= 0))
536 "Can not set nstlist without %s",
537 !EI_DYNAMICS(ir->eI) ? "dynamics" : "verlet-buffer-tolerance");
540 if (EI_DYNAMICS(ir->eI))
542 /* Set or try nstlist values */
543 increaseNstlist(fplog, cr, ir, nstlist_cmdline, &mtop, box, makeGpuPairList, cpuinfo);
547 /*! \brief Override the nslist value in inputrec
549 * with value passed on the command line (if any)
551 static void override_nsteps_cmdline(const gmx::MDLogger& mdlog, int64_t nsteps_cmdline, t_inputrec* ir)
555 /* override with anything else than the default -2 */
556 if (nsteps_cmdline > -2)
558 char sbuf_steps[STEPSTRSIZE];
559 char sbuf_msg[STRLEN];
561 ir->nsteps = nsteps_cmdline;
562 if (EI_DYNAMICS(ir->eI) && nsteps_cmdline != -1)
565 "Overriding nsteps with value passed on the command line: %s steps, %.3g ps",
566 gmx_step_str(nsteps_cmdline, sbuf_steps),
567 fabs(nsteps_cmdline * ir->delta_t));
572 "Overriding nsteps with value passed on the command line: %s steps",
573 gmx_step_str(nsteps_cmdline, sbuf_steps));
576 GMX_LOG(mdlog.warning).asParagraph().appendText(sbuf_msg);
578 else if (nsteps_cmdline < -2)
580 gmx_fatal(FARGS, "Invalid nsteps value passed on the command line: %" PRId64, nsteps_cmdline);
582 /* Do nothing if nsteps_cmdline == -2 */
588 /*! \brief Return whether GPU acceleration of nonbondeds is supported with the given settings.
590 * If not, and if a warning may be issued, logs a warning about
591 * falling back to CPU code. With thread-MPI, only the first
592 * call to this function should have \c issueWarning true. */
593 static bool gpuAccelerationOfNonbondedIsUseful(const MDLogger& mdlog, const t_inputrec& ir, bool issueWarning)
595 bool gpuIsUseful = true;
598 if (ir.opts.ngener - ir.nwall > 1)
600 /* The GPU code does not support more than one energy group.
601 * If the user requested GPUs explicitly, a fatal error is given later.
605 "Multiple energy groups is not implemented for GPUs, falling back to the CPU. "
606 "For better performance, run on the GPU without energy groups and then do "
607 "gmx mdrun -rerun option on the trajectory with an energy group .tpr file.";
613 warning = "TPI is not implemented for GPUs.";
616 if (!gpuIsUseful && issueWarning)
618 GMX_LOG(mdlog.warning).asParagraph().appendText(warning);
624 //! Initializes the logger for mdrun.
625 static gmx::LoggerOwner buildLogger(FILE* fplog, const bool isSimulationMasterRank)
627 gmx::LoggerBuilder builder;
628 if (fplog != nullptr)
630 builder.addTargetFile(gmx::MDLogger::LogLevel::Info, fplog);
632 if (isSimulationMasterRank)
634 builder.addTargetStream(gmx::MDLogger::LogLevel::Warning, &gmx::TextOutputFile::standardError());
636 return builder.build();
639 //! Make a TaskTarget from an mdrun argument string.
640 static TaskTarget findTaskTarget(const char* optionString)
642 TaskTarget returnValue = TaskTarget::Auto;
644 if (strncmp(optionString, "auto", 3) == 0)
646 returnValue = TaskTarget::Auto;
648 else if (strncmp(optionString, "cpu", 3) == 0)
650 returnValue = TaskTarget::Cpu;
652 else if (strncmp(optionString, "gpu", 3) == 0)
654 returnValue = TaskTarget::Gpu;
658 GMX_ASSERT(false, "Option string should have been checked for sanity already");
664 //! Finish run, aggregate data to print performance info.
665 static void finish_run(FILE* fplog,
666 const gmx::MDLogger& mdlog,
668 const t_inputrec& inputrec,
670 gmx_wallcycle* wcycle,
671 gmx_walltime_accounting_t walltime_accounting,
672 nonbonded_verlet_t* nbv,
673 const gmx_pme_t* pme,
677 double nbfs = 0, mflop = 0;
678 double elapsed_time, elapsed_time_over_all_ranks, elapsed_time_over_all_threads,
679 elapsed_time_over_all_threads_over_all_ranks;
680 /* Control whether it is valid to print a report. Only the
681 simulation master may print, but it should not do so if the run
682 terminated e.g. before a scheduled reset step. This is
683 complicated by the fact that PME ranks are unaware of the
684 reason why they were sent a pmerecvqxFINISH. To avoid
685 communication deadlocks, we always do the communication for the
686 report, even if we've decided not to write the report, because
687 how long it takes to finish the run is not important when we've
688 decided not to report on the simulation performance.
690 Further, we only report performance for dynamical integrators,
691 because those are the only ones for which we plan to
692 consider doing any optimizations. */
693 bool printReport = EI_DYNAMICS(inputrec.eI) && SIMMASTER(cr);
695 if (printReport && !walltime_accounting_get_valid_finish(walltime_accounting))
697 GMX_LOG(mdlog.warning)
699 .appendText("Simulation ended prematurely, no performance report will be written.");
704 std::unique_ptr<t_nrnb> nrnbTotalStorage;
707 nrnbTotalStorage = std::make_unique<t_nrnb>();
708 nrnb_tot = nrnbTotalStorage.get();
710 MPI_Allreduce(nrnb->n.data(), nrnb_tot->n.data(), eNRNB, MPI_DOUBLE, MPI_SUM, cr->mpi_comm_mysim);
718 elapsed_time = walltime_accounting_get_time_since_reset(walltime_accounting);
719 elapsed_time_over_all_threads =
720 walltime_accounting_get_time_since_reset_over_all_threads(walltime_accounting);
724 /* reduce elapsed_time over all MPI ranks in the current simulation */
725 MPI_Allreduce(&elapsed_time, &elapsed_time_over_all_ranks, 1, MPI_DOUBLE, MPI_SUM, cr->mpi_comm_mysim);
726 elapsed_time_over_all_ranks /= cr->nnodes;
727 /* Reduce elapsed_time_over_all_threads over all MPI ranks in the
728 * current simulation. */
729 MPI_Allreduce(&elapsed_time_over_all_threads,
730 &elapsed_time_over_all_threads_over_all_ranks,
739 elapsed_time_over_all_ranks = elapsed_time;
740 elapsed_time_over_all_threads_over_all_ranks = elapsed_time_over_all_threads;
745 print_flop(fplog, nrnb_tot, &nbfs, &mflop);
748 if (thisRankHasDuty(cr, DUTY_PP) && DOMAINDECOMP(cr))
750 print_dd_statistics(cr, inputrec, fplog);
753 /* TODO Move the responsibility for any scaling by thread counts
754 * to the code that handled the thread region, so that there's a
755 * mechanism to keep cycle counting working during the transition
756 * to task parallelism. */
757 int nthreads_pp = gmx_omp_nthreads_get(ModuleMultiThread::Nonbonded);
758 int nthreads_pme = gmx_omp_nthreads_get(ModuleMultiThread::Pme);
759 wallcycle_scale_by_num_threads(
760 wcycle, thisRankHasDuty(cr, DUTY_PME) && !thisRankHasDuty(cr, DUTY_PP), nthreads_pp, nthreads_pme);
761 auto cycle_sum(wallcycle_sum(cr, wcycle));
765 auto* nbnxn_gpu_timings =
766 (nbv != nullptr && nbv->useGpu()) ? Nbnxm::gpu_get_timings(nbv->gpu_nbv) : nullptr;
767 gmx_wallclock_gpu_pme_t pme_gpu_timings = {};
769 if (pme_gpu_task_enabled(pme))
771 pme_gpu_get_timings(pme, &pme_gpu_timings);
773 wallcycle_print(fplog,
779 elapsed_time_over_all_ranks,
785 if (EI_DYNAMICS(inputrec.eI))
787 delta_t = inputrec.delta_t;
793 elapsed_time_over_all_threads_over_all_ranks,
794 elapsed_time_over_all_ranks,
795 walltime_accounting_get_nsteps_done_since_reset(walltime_accounting),
803 elapsed_time_over_all_threads_over_all_ranks,
804 elapsed_time_over_all_ranks,
805 walltime_accounting_get_nsteps_done_since_reset(walltime_accounting),
813 int Mdrunner::mdrunner()
816 std::unique_ptr<t_forcerec> fr;
817 real ewaldcoeff_q = 0;
818 real ewaldcoeff_lj = 0;
819 int nChargePerturbed = -1, nTypePerturbed = 0;
820 gmx_walltime_accounting_t walltime_accounting = nullptr;
821 MembedHolder membedHolder(filenames.size(), filenames.data());
823 /* CAUTION: threads may be started later on in this function, so
824 cr doesn't reflect the final parallel state right now */
827 /* TODO: inputrec should tell us whether we use an algorithm, not a file option */
828 const bool doEssentialDynamics = opt2bSet("-ei", filenames.size(), filenames.data());
829 const bool doRerun = mdrunOptions.rerun;
831 // Handle task-assignment related user options.
832 EmulateGpuNonbonded emulateGpuNonbonded =
833 (getenv("GMX_EMULATE_GPU") != nullptr ? EmulateGpuNonbonded::Yes : EmulateGpuNonbonded::No);
835 std::vector<int> userGpuTaskAssignment;
838 userGpuTaskAssignment = parseUserTaskAssignmentString(hw_opt.userGpuTaskAssignment);
840 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
841 auto nonbondedTarget = findTaskTarget(nbpu_opt);
842 auto pmeTarget = findTaskTarget(pme_opt);
843 auto pmeFftTarget = findTaskTarget(pme_fft_opt);
844 auto bondedTarget = findTaskTarget(bonded_opt);
845 auto updateTarget = findTaskTarget(update_opt);
847 FILE* fplog = nullptr;
848 // If we are appending, we don't write log output because we need
849 // to check that the old log file matches what the checkpoint file
850 // expects. Otherwise, we should start to write log output now if
851 // there is a file ready for it.
852 if (logFileHandle != nullptr && startingBehavior != StartingBehavior::RestartWithAppending)
854 fplog = gmx_fio_getfp(logFileHandle);
856 const bool isSimulationMasterRank = findIsSimulationMasterRank(ms, simulationCommunicator);
857 gmx::LoggerOwner logOwner(buildLogger(fplog, isSimulationMasterRank));
858 gmx::MDLogger mdlog(logOwner.logger());
860 gmx_print_detected_hardware(fplog, isSimulationMasterRank && isMasterSim(ms), mdlog, hwinfo_);
862 std::vector<int> availableDevices =
863 makeListOfAvailableDevices(hwinfo_->deviceInfoList, hw_opt.devicesSelectedByUser);
864 const int numAvailableDevices = gmx::ssize(availableDevices);
866 // Print citation requests after all software/hardware printing
867 pleaseCiteGromacs(fplog);
869 // Note: legacy program logic relies on checking whether these pointers are assigned.
870 // Objects may or may not be allocated later.
871 std::unique_ptr<t_inputrec> inputrec;
872 std::unique_ptr<t_state> globalState;
874 auto partialDeserializedTpr = std::make_unique<PartialDeserializedTprFile>();
876 if (isSimulationMasterRank)
878 // Allocate objects to be initialized by later function calls.
879 /* Only the master rank has the global state */
880 globalState = std::make_unique<t_state>();
881 inputrec = std::make_unique<t_inputrec>();
883 /* Read (nearly) all data required for the simulation
884 * and keep the partly serialized tpr contents to send to other ranks later
886 applyGlobalSimulationState(
887 *inputHolder_.get(), partialDeserializedTpr.get(), globalState.get(), inputrec.get(), &mtop);
890 /* Check and update the hardware options for internal consistency */
891 checkAndUpdateHardwareOptions(
892 mdlog, &hw_opt, isSimulationMasterRank, domdecOptions.numPmeRanks, inputrec.get());
894 if (GMX_THREAD_MPI && isSimulationMasterRank)
896 bool useGpuForNonbonded = false;
897 bool useGpuForPme = false;
900 GMX_RELEASE_ASSERT(inputrec != nullptr, "Keep the compiler happy");
902 // If the user specified the number of ranks, then we must
903 // respect that, but in default mode, we need to allow for
904 // the number of GPUs to choose the number of ranks.
905 auto canUseGpuForNonbonded = buildSupportsNonbondedOnGpu(nullptr);
906 useGpuForNonbonded = decideWhetherToUseGpusForNonbondedWithThreadMpi(
908 numAvailableDevices > 0,
909 userGpuTaskAssignment,
911 canUseGpuForNonbonded,
912 gpuAccelerationOfNonbondedIsUseful(mdlog, *inputrec, GMX_THREAD_MPI),
913 hw_opt.nthreads_tmpi);
914 useGpuForPme = decideWhetherToUseGpusForPmeWithThreadMpi(useGpuForNonbonded,
917 userGpuTaskAssignment,
920 hw_opt.nthreads_tmpi,
921 domdecOptions.numPmeRanks);
923 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
925 /* Determine how many thread-MPI ranks to start.
927 * TODO Over-writing the user-supplied value here does
928 * prevent any possible subsequent checks from working
930 hw_opt.nthreads_tmpi = get_nthreads_mpi(hwinfo_,
938 membedHolder.doMembed());
940 // Now start the threads for thread MPI.
941 spawnThreads(hw_opt.nthreads_tmpi);
942 // The spawned threads enter mdrunner() and execution of
943 // master and spawned threads joins at the end of this block.
946 GMX_RELEASE_ASSERT(ms || simulationCommunicator != MPI_COMM_NULL,
947 "Must have valid communicator unless running a multi-simulation");
948 CommrecHandle crHandle = init_commrec(simulationCommunicator);
949 t_commrec* cr = crHandle.get();
950 GMX_RELEASE_ASSERT(cr != nullptr, "Must have valid commrec");
952 PhysicalNodeCommunicator physicalNodeComm(libraryWorldCommunicator, gmx_physicalnode_id_hash());
954 // If we detected the topology on this system, double-check that it makes sense
955 if (hwinfo_->hardwareTopology->isThisSystem())
957 hardwareTopologyDoubleCheckDetection(mdlog, *hwinfo_->hardwareTopology);
962 /* now broadcast everything to the non-master nodes/threads: */
963 if (!isSimulationMasterRank)
965 // Until now, only the master rank has a non-null pointer.
966 // On non-master ranks, allocate the object that will receive data in the following call.
967 inputrec = std::make_unique<t_inputrec>();
969 init_parallel(cr->mpiDefaultCommunicator,
973 partialDeserializedTpr.get());
975 GMX_RELEASE_ASSERT(inputrec != nullptr, "All ranks should have a valid inputrec now");
976 partialDeserializedTpr.reset(nullptr);
979 !inputrec->useConstantAcceleration,
980 "Linear acceleration has been removed in GROMACS 2022, and was broken for many years "
981 "before that. Use GROMACS 4.5 or earlier if you need this feature.");
983 // Now the number of ranks is known to all ranks, and each knows
984 // the inputrec read by the master rank. The ranks can now all run
985 // the task-deciding functions and will agree on the result
986 // without needing to communicate.
987 const bool useDomainDecomposition =
988 (PAR(cr) && !(EI_TPI(inputrec->eI) || inputrec->eI == IntegrationAlgorithm::NM));
990 // Note that these variables describe only their own node.
992 // Note that when bonded interactions run on a GPU they always run
993 // alongside a nonbonded task, so do not influence task assignment
994 // even though they affect the force calculation workload.
995 bool useGpuForNonbonded = false;
996 bool useGpuForPme = false;
997 bool useGpuForBonded = false;
998 bool useGpuForUpdate = false;
999 bool gpusWereDetected = hwinfo_->ngpu_compatible_tot > 0;
1002 // It's possible that there are different numbers of GPUs on
1003 // different nodes, which is the user's responsibility to
1004 // handle. If unsuitable, we will notice that during task
1006 auto canUseGpuForNonbonded = buildSupportsNonbondedOnGpu(nullptr);
1007 useGpuForNonbonded = decideWhetherToUseGpusForNonbonded(
1009 userGpuTaskAssignment,
1010 emulateGpuNonbonded,
1011 canUseGpuForNonbonded,
1012 gpuAccelerationOfNonbondedIsUseful(mdlog, *inputrec, !GMX_THREAD_MPI),
1014 useGpuForPme = decideWhetherToUseGpusForPme(useGpuForNonbonded,
1016 userGpuTaskAssignment,
1019 cr->sizeOfDefaultCommunicator,
1020 domdecOptions.numPmeRanks,
1022 useGpuForBonded = decideWhetherToUseGpusForBonded(
1023 useGpuForNonbonded, useGpuForPme, bondedTarget, *inputrec, mtop, domdecOptions.numPmeRanks, gpusWereDetected);
1025 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1027 const PmeRunMode pmeRunMode = determinePmeRunMode(useGpuForPme, pmeFftTarget, *inputrec);
1029 // Initialize development feature flags that enabled by environment variable
1030 // and report those features that are enabled.
1031 const DevelopmentFeatureFlags devFlags =
1032 manageDevelopmentFeatures(mdlog, useGpuForNonbonded, pmeRunMode);
1034 const bool useModularSimulator = checkUseModularSimulator(false,
1041 doEssentialDynamics,
1042 membedHolder.doMembed());
1044 // Build restraints.
1045 // TODO: hide restraint implementation details from Mdrunner.
1046 // There is nothing unique about restraints at this point as far as the
1047 // Mdrunner is concerned. The Mdrunner should just be getting a sequence of
1048 // factory functions from the SimulationContext on which to call mdModules_->add().
1049 // TODO: capture all restraints into a single RestraintModule, passed to the runner builder.
1050 for (auto&& restraint : restraintManager_->getRestraints())
1052 auto module = RestraintMDModule::create(restraint, restraint->sites());
1053 mdModules_->add(std::move(module));
1056 // TODO: Error handling
1057 mdModules_->assignOptionsToModules(*inputrec->params, nullptr);
1058 // now that the MDModules know their options, they know which callbacks to sign up to
1059 mdModules_->subscribeToSimulationSetupNotifications();
1060 const auto& setupNotifier = mdModules_->notifiers().simulationSetupNotifier_;
1062 // Notify MdModules of existing logger
1063 setupNotifier.notify(mdlog);
1065 // Notify MdModules of internal parameters, saved into KVT
1066 if (inputrec->internalParameters != nullptr)
1068 setupNotifier.notify(*inputrec->internalParameters);
1071 // Let MdModules know the .tpr filename
1073 gmx::MdRunInputFilename mdRunInputFilename = { ftp2fn(efTPR, filenames.size(), filenames.data()) };
1074 setupNotifier.notify(mdRunInputFilename);
1077 if (fplog != nullptr)
1079 pr_inputrec(fplog, 0, "Input Parameters", inputrec.get(), FALSE);
1080 fprintf(fplog, "\n");
1085 /* In rerun, set velocities to zero if present */
1086 if (doRerun && ((globalState->flags & enumValueToBitMask(StateEntry::V)) != 0))
1088 // rerun does not use velocities
1092 "Rerun trajectory contains velocities. Rerun does only evaluate "
1093 "potential energy and forces. The velocities will be ignored.");
1094 for (int i = 0; i < globalState->natoms; i++)
1096 clear_rvec(globalState->v[i]);
1098 globalState->flags &= ~enumValueToBitMask(StateEntry::V);
1101 /* now make sure the state is initialized and propagated */
1102 set_state_entries(globalState.get(), inputrec.get(), useModularSimulator);
1105 /* NM and TPI parallelize over force/energy calculations, not atoms,
1106 * so we need to initialize and broadcast the global state.
1108 if (inputrec->eI == IntegrationAlgorithm::NM || inputrec->eI == IntegrationAlgorithm::TPI)
1112 globalState = std::make_unique<t_state>();
1114 broadcastStateWithoutDynamics(
1115 cr->mpiDefaultCommunicator, DOMAINDECOMP(cr), PAR(cr), globalState.get());
1118 /* A parallel command line option consistency check that we can
1119 only do after any threads have started. */
1121 && (domdecOptions.numCells[XX] > 1 || domdecOptions.numCells[YY] > 1
1122 || domdecOptions.numCells[ZZ] > 1 || domdecOptions.numPmeRanks > 0))
1125 "The -dd or -npme option request a parallel simulation, "
1127 "but %s was compiled without threads or MPI enabled",
1128 output_env_get_program_display_name(oenv));
1129 #elif GMX_THREAD_MPI
1130 "but the number of MPI-threads (option -ntmpi) is not set or is 1");
1132 "but %s was not started through mpirun/mpiexec or only one rank was requested "
1133 "through mpirun/mpiexec",
1134 output_env_get_program_display_name(oenv));
1138 if (doRerun && (EI_ENERGY_MINIMIZATION(inputrec->eI) || IntegrationAlgorithm::NM == inputrec->eI))
1141 "The .mdp file specified an energy mininization or normal mode algorithm, and "
1142 "these are not compatible with mdrun -rerun");
1145 /* NMR restraints must be initialized before load_checkpoint,
1146 * since with time averaging the history is added to t_state.
1147 * For proper consistency check we therefore need to extend
1149 * So the PME-only nodes (if present) will also initialize
1150 * the distance restraints.
1153 /* This needs to be called before read_checkpoint to extend the state */
1154 t_disresdata* disresdata;
1155 snew(disresdata, 1);
1159 DisResRunMode::MDRun,
1160 MASTER(cr) ? DDRole::Master : DDRole::Agent,
1161 PAR(cr) ? NumRanks::Multiple : NumRanks::Single,
1166 replExParams.exchangeInterval > 0);
1168 std::unique_ptr<t_oriresdata> oriresData;
1169 if (gmx_mtop_ftype_count(mtop, F_ORIRES) > 0)
1171 oriresData = std::make_unique<t_oriresdata>(fplog, mtop, *inputrec, cr, ms, globalState.get());
1174 auto deform = prepareBoxDeformation(globalState != nullptr ? globalState->box : box,
1175 MASTER(cr) ? DDRole::Master : DDRole::Agent,
1176 PAR(cr) ? NumRanks::Multiple : NumRanks::Single,
1177 cr->mpi_comm_mygroup,
1181 /* We have to remember the generation's first step before reading checkpoint.
1182 This way, we can report to the F@H core both the generation's first step
1183 and the restored first step, thus making it able to distinguish between
1184 an interruption/resume and start of the n-th generation simulation.
1185 Having this information, the F@H core can correctly calculate and report
1188 int gen_first_step = 0;
1191 gen_first_step = inputrec->init_step;
1195 ObservablesHistory observablesHistory = {};
1197 auto modularSimulatorCheckpointData = std::make_unique<ReadCheckpointDataHolder>();
1198 if (startingBehavior != StartingBehavior::NewSimulation)
1200 /* Check if checkpoint file exists before doing continuation.
1201 * This way we can use identical input options for the first and subsequent runs...
1203 if (mdrunOptions.numStepsCommandline > -2)
1205 /* Temporarily set the number of steps to unlimited to avoid
1206 * triggering the nsteps check in load_checkpoint().
1207 * This hack will go away soon when the -nsteps option is removed.
1209 inputrec->nsteps = -1;
1212 // Finish applying initial simulation state information from external sources on all ranks.
1213 // Reconcile checkpoint file data with Mdrunner state established up to this point.
1214 applyLocalState(*inputHolder_.get(),
1217 domdecOptions.numCells,
1220 &observablesHistory,
1221 mdrunOptions.reproducible,
1222 mdModules_->notifiers(),
1223 modularSimulatorCheckpointData.get(),
1224 useModularSimulator);
1225 // TODO: (#3652) Synchronize filesystem state, SimulationInput contents, and program
1227 // on all code paths.
1228 // Write checkpoint or provide hook to update SimulationInput.
1229 // If there was a checkpoint file, SimulationInput contains more information
1230 // than if there wasn't. At this point, we have synchronized the in-memory
1231 // state with the filesystem state only for restarted simulations. We should
1232 // be calling applyLocalState unconditionally and expect that the completeness
1233 // of SimulationInput is not dependent on its creation method.
1235 if (startingBehavior == StartingBehavior::RestartWithAppending && logFileHandle)
1237 // Now we can start normal logging to the truncated log file.
1238 fplog = gmx_fio_getfp(logFileHandle);
1239 prepareLogAppending(fplog);
1240 logOwner = buildLogger(fplog, MASTER(cr));
1241 mdlog = logOwner.logger();
1248 fcRegisterSteps(inputrec->nsteps + inputrec->init_step, gen_first_step);
1252 if (mdrunOptions.numStepsCommandline > -2)
1257 "The -nsteps functionality is deprecated, and may be removed in a future "
1259 "Consider using gmx convert-tpr -nsteps or changing the appropriate .mdp "
1262 /* override nsteps with value set on the commandline */
1263 override_nsteps_cmdline(mdlog, mdrunOptions.numStepsCommandline, inputrec.get());
1265 if (isSimulationMasterRank)
1267 copy_mat(globalState->box, box);
1272 gmx_bcast(sizeof(box), box, cr->mpiDefaultCommunicator);
1275 if (inputrec->cutoff_scheme != CutoffScheme::Verlet)
1278 "This group-scheme .tpr file can no longer be run by mdrun. Please update to the "
1279 "Verlet scheme, or use an earlier version of GROMACS if necessary.");
1281 /* Update rlist and nstlist. */
1282 /* Note: prepare_verlet_scheme is calling increaseNstlist(...), which (while attempting to
1283 * increase rlist) tries to check if the newly chosen value fits with the DD scheme. As this is
1284 * run before any DD scheme is set up, this check is never executed. See #3334 for more details.
1286 prepare_verlet_scheme(fplog,
1292 useGpuForNonbonded || (emulateGpuNonbonded == EmulateGpuNonbonded::Yes),
1295 // We need to decide on update groups early, as this affects
1296 // inter-domain communication distances.
1297 auto updateGroupingsPerMoleculeType = makeUpdateGroupingsPerMoleculeType(mtop);
1298 const real maxUpdateGroupRadius = computeMaxUpdateGroupRadius(
1299 mtop, updateGroupingsPerMoleculeType, maxReferenceTemperature(*inputrec));
1300 const real cutoffMargin = std::sqrt(max_cutoff2(inputrec->pbcType, box)) - inputrec->rlist;
1301 UpdateGroups updateGroups = makeUpdateGroups(mdlog,
1302 std::move(updateGroupingsPerMoleculeType),
1303 maxUpdateGroupRadius,
1304 useDomainDecomposition,
1305 systemHasConstraintsOrVsites(mtop),
1308 // This builder is necessary while we have multi-part construction
1309 // of DD. Before DD is constructed, we use the existence of
1310 // the builder object to indicate that further construction of DD
1312 std::unique_ptr<DomainDecompositionBuilder> ddBuilder;
1313 if (useDomainDecomposition)
1315 ddBuilder = std::make_unique<DomainDecompositionBuilder>(
1322 mdModules_->notifiers(),
1324 updateGroups.updateGroupingPerMoleculeType(),
1325 updateGroups.useUpdateGroups(),
1326 updateGroups.maxUpdateGroupRadius(),
1327 positionsFromStatePointer(globalState.get()),
1333 /* PME, if used, is done on all nodes with 1D decomposition */
1334 cr->nnodes = cr->sizeOfDefaultCommunicator;
1335 cr->sim_nodeid = cr->rankInDefaultCommunicator;
1336 cr->nodeid = cr->rankInDefaultCommunicator;
1338 cr->duty = (DUTY_PP | DUTY_PME);
1340 if (inputrec->pbcType == PbcType::Screw)
1342 gmx_fatal(FARGS, "pbc=screw is only implemented with domain decomposition");
1346 // Produce the task assignment for this rank - done after DD is constructed
1347 GpuTaskAssignments gpuTaskAssignments = GpuTaskAssignmentsBuilder::build(
1349 userGpuTaskAssignment,
1351 simulationCommunicator,
1359 thisRankHasDuty(cr, DUTY_PP),
1360 // TODO cr->duty & DUTY_PME should imply that a PME
1361 // algorithm is active, but currently does not.
1362 EEL_PME(inputrec->coulombtype) && thisRankHasDuty(cr, DUTY_PME));
1364 // Get the device handles for the modules, nullptr when no task is assigned.
1366 DeviceInformation* deviceInfo = gpuTaskAssignments.initDevice(&deviceId);
1368 // timing enabling - TODO put this in gpu_utils (even though generally this is just option handling?)
1369 bool useTiming = true;
1373 /* WARNING: CUDA timings are incorrect with multiple streams.
1374 * This is the main reason why they are disabled by default.
1376 // TODO: Consider turning on by default when we can detect nr of streams.
1377 useTiming = (getenv("GMX_ENABLE_GPU_TIMING") != nullptr);
1379 else if (GMX_GPU_OPENCL)
1381 useTiming = (getenv("GMX_DISABLE_GPU_TIMING") == nullptr);
1384 // TODO Currently this is always built, yet DD partition code
1385 // checks if it is built before using it. Probably it should
1386 // become an MDModule that is made only when another module
1387 // requires it (e.g. pull, CompEl, density fitting), so that we
1388 // don't update the local atom sets unilaterally every step.
1389 LocalAtomSetManager atomSets;
1392 // TODO Pass the GPU streams to ddBuilder to use in buffer
1393 // transfers (e.g. halo exchange)
1394 cr->dd = ddBuilder->build(&atomSets);
1395 // The builder's job is done, so destruct it
1396 ddBuilder.reset(nullptr);
1397 // Note that local state still does not exist yet.
1400 // The GPU update is decided here because we need to know whether the constraints or
1401 // SETTLEs can span accross the domain borders (i.e. whether or not update groups are
1402 // defined). This is only known after DD is initialized, hence decision on using GPU
1403 // update is done so late.
1406 const bool haveFrozenAtoms = inputrecFrozenAtoms(inputrec.get());
1408 useGpuForUpdate = decideWhetherToUseGpuForUpdate(useDomainDecomposition,
1409 updateGroups.useUpdateGroups(),
1411 domdecOptions.numPmeRanks > 0,
1417 doEssentialDynamics,
1418 gmx_mtop_ftype_count(mtop, F_ORIRES) > 0,
1424 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1426 const bool printHostName = (cr->nnodes > 1);
1427 gpuTaskAssignments.reportGpuUsage(mdlog, printHostName, useGpuForBonded, pmeRunMode, useGpuForUpdate);
1429 const bool disableNonbondedCalculation = (getenv("GMX_NO_NONBONDED") != nullptr);
1430 if (disableNonbondedCalculation)
1432 /* turn off non-bonded calculations */
1433 GMX_LOG(mdlog.warning)
1436 "Found environment variable GMX_NO_NONBONDED.\n"
1437 "Disabling nonbonded calculations.");
1440 MdrunScheduleWorkload runScheduleWork;
1442 bool useGpuDirectHalo = decideWhetherToUseGpuForHalo(devFlags,
1443 havePPDomainDecomposition(cr),
1445 useModularSimulator,
1447 EI_ENERGY_MINIMIZATION(inputrec->eI));
1449 // Also populates the simulation constant workload description.
1450 runScheduleWork.simulationWork = createSimulationWorkload(*inputrec,
1451 disableNonbondedCalculation,
1459 std::unique_ptr<DeviceStreamManager> deviceStreamManager = nullptr;
1461 if (deviceInfo != nullptr)
1463 if (DOMAINDECOMP(cr) && thisRankHasDuty(cr, DUTY_PP))
1465 dd_setup_dlb_resource_sharing(cr, deviceId);
1467 deviceStreamManager = std::make_unique<DeviceStreamManager>(
1468 *deviceInfo, havePPDomainDecomposition(cr), runScheduleWork.simulationWork, useTiming);
1471 // If the user chose a task assignment, give them some hints
1472 // where appropriate.
1473 if (!userGpuTaskAssignment.empty())
1475 gpuTaskAssignments.logPerformanceHints(mdlog, numAvailableDevices);
1480 /* After possible communicator splitting in make_dd_communicators.
1481 * we can set up the intra/inter node communication.
1483 gmx_setup_nodecomm(fplog, cr);
1489 GMX_LOG(mdlog.warning)
1491 .appendTextFormatted(
1492 "This is simulation %d out of %d running as a composite GROMACS\n"
1493 "multi-simulation job. Setup for this simulation:\n",
1494 ms->simulationIndex_,
1495 ms->numSimulations_);
1497 GMX_LOG(mdlog.warning)
1498 .appendTextFormatted("Using %d MPI %s\n",
1501 cr->nnodes == 1 ? "thread" : "threads"
1503 cr->nnodes == 1 ? "process" : "processes"
1509 // If mdrun -pin auto honors any affinity setting that already
1510 // exists. If so, it is nice to provide feedback about whether
1511 // that existing affinity setting was from OpenMP or something
1512 // else, so we run this code both before and after we initialize
1513 // the OpenMP support.
1514 gmx_check_thread_affinity_set(mdlog, &hw_opt, hwinfo_->nthreads_hw_avail, FALSE);
1515 /* Check and update the number of OpenMP threads requested */
1516 checkAndUpdateRequestedNumOpenmpThreads(
1517 &hw_opt, *hwinfo_, cr, ms, physicalNodeComm.size_, pmeRunMode, mtop, *inputrec);
1519 gmx_omp_nthreads_init(mdlog,
1521 hwinfo_->nthreads_hw_avail,
1522 physicalNodeComm.size_,
1523 hw_opt.nthreads_omp,
1524 hw_opt.nthreads_omp_pme,
1525 !thisRankHasDuty(cr, DUTY_PP));
1527 const bool bEnableFPE = gmxShouldEnableFPExceptions();
1528 // FIXME - reconcile with gmx_feenableexcept() call from CommandLineModuleManager::run()
1531 gmx_feenableexcept();
1534 /* Now that we know the setup is consistent, check for efficiency */
1535 check_resource_division_efficiency(
1536 hwinfo_, gpuTaskAssignments.thisRankHasAnyGpuTask(), mdrunOptions.ntompOptionIsSet, cr, mdlog);
1538 /* getting number of PP/PME threads on this MPI / tMPI rank.
1539 PME: env variable should be read only on one node to make sure it is
1540 identical everywhere;
1542 const int numThreadsOnThisRank = thisRankHasDuty(cr, DUTY_PP)
1543 ? gmx_omp_nthreads_get(ModuleMultiThread::Nonbonded)
1544 : gmx_omp_nthreads_get(ModuleMultiThread::Pme);
1545 checkHardwareOversubscription(
1546 numThreadsOnThisRank, cr->nodeid, *hwinfo_->hardwareTopology, physicalNodeComm, mdlog);
1548 // Enable Peer access between GPUs where available
1549 // Only for DD, only master PP rank needs to perform setup, and only if thread MPI plus
1550 // any of the GPU communication features are active.
1551 if (DOMAINDECOMP(cr) && MASTER(cr) && thisRankHasDuty(cr, DUTY_PP) && GMX_THREAD_MPI
1552 && (runScheduleWork.simulationWork.useGpuHaloExchange
1553 || runScheduleWork.simulationWork.useGpuPmePpCommunication))
1555 setupGpuDevicePeerAccess(gpuTaskAssignments.deviceIdsAssigned(), mdlog);
1558 if (hw_opt.threadAffinity != ThreadAffinity::Off)
1560 /* Before setting affinity, check whether the affinity has changed
1561 * - which indicates that probably the OpenMP library has changed it
1562 * since we first checked).
1564 gmx_check_thread_affinity_set(mdlog, &hw_opt, hwinfo_->nthreads_hw_avail, TRUE);
1566 int numThreadsOnThisNode, intraNodeThreadOffset;
1567 analyzeThreadsOnThisNode(
1568 physicalNodeComm, numThreadsOnThisRank, &numThreadsOnThisNode, &intraNodeThreadOffset);
1570 /* Set the CPU affinity */
1571 gmx_set_thread_affinity(mdlog,
1574 *hwinfo_->hardwareTopology,
1575 numThreadsOnThisRank,
1576 numThreadsOnThisNode,
1577 intraNodeThreadOffset,
1581 if (mdrunOptions.timingOptions.resetStep > -1)
1586 "The -resetstep functionality is deprecated, and may be removed in a "
1589 std::unique_ptr<gmx_wallcycle> wcycle =
1590 wallcycle_init(fplog, mdrunOptions.timingOptions.resetStep, cr);
1594 /* Master synchronizes its value of reset_counters with all nodes
1595 * including PME only nodes */
1596 int64_t reset_counters = wcycle_get_reset_counters(wcycle.get());
1597 gmx_bcast(sizeof(reset_counters), &reset_counters, cr->mpi_comm_mysim);
1598 wcycle_set_reset_counters(wcycle.get(), reset_counters);
1601 // Membrane embedding must be initialized before we call init_forcerec()
1602 membedHolder.initializeMembed(fplog,
1609 &mdrunOptions.checkpointOptions.period);
1611 const bool thisRankHasPmeGpuTask = gpuTaskAssignments.thisRankHasPmeGpuTask();
1612 std::unique_ptr<MDAtoms> mdAtoms;
1613 std::unique_ptr<VirtualSitesHandler> vsite;
1614 std::unique_ptr<ListedForcesGpu> listedForcesGpu;
1617 if (thisRankHasDuty(cr, DUTY_PP))
1619 setupNotifier.notify(*cr);
1620 setupNotifier.notify(&atomSets);
1621 setupNotifier.notify(mtop);
1622 setupNotifier.notify(inputrec->pbcType);
1623 setupNotifier.notify(SimulationTimeStep{ inputrec->delta_t });
1624 /* Initiate forcerecord */
1625 fr = std::make_unique<t_forcerec>();
1626 fr->forceProviders = mdModules_->initForceProviders();
1627 init_forcerec(fplog,
1629 runScheduleWork.simulationWork,
1635 opt2fn("-table", filenames.size(), filenames.data()),
1636 opt2fn("-tablep", filenames.size(), filenames.data()),
1637 opt2fns("-tableb", filenames.size(), filenames.data()),
1639 // Dirty hack, for fixing disres and orires should be made mdmodules
1640 fr->fcdata->disres = disresdata;
1641 fr->fcdata->orires.swap(oriresData);
1643 // Save a handle to device stream manager to use elsewhere in the code
1644 // TODO: Forcerec is not a correct place to store it.
1645 fr->deviceStreamManager = deviceStreamManager.get();
1647 if (runScheduleWork.simulationWork.useGpuPmePpCommunication && !thisRankHasDuty(cr, DUTY_PME))
1650 deviceStreamManager != nullptr,
1651 "GPU device stream manager should be valid in order to use PME-PP direct "
1654 deviceStreamManager->streamIsValid(DeviceStreamType::PmePpTransfer),
1655 "GPU PP-PME stream should be valid in order to use GPU PME-PP direct "
1657 fr->pmePpCommGpu = std::make_unique<gmx::PmePpCommGpu>(
1660 deviceStreamManager->context(),
1661 deviceStreamManager->stream(DeviceStreamType::PmePpTransfer));
1664 fr->nbv = Nbnxm::init_nb_verlet(mdlog,
1669 runScheduleWork.simulationWork.useGpuNonbonded,
1670 deviceStreamManager.get(),
1674 // TODO: Move the logic below to a GPU bonded builder
1675 if (runScheduleWork.simulationWork.useGpuBonded)
1677 GMX_RELEASE_ASSERT(deviceStreamManager != nullptr,
1678 "GPU device stream manager should be valid in order to use GPU "
1679 "version of bonded forces.");
1680 listedForcesGpu = std::make_unique<ListedForcesGpu>(
1682 fr->ic->epsfac * fr->fudgeQQ,
1683 deviceStreamManager->context(),
1684 deviceStreamManager->bondedStream(havePPDomainDecomposition(cr)),
1686 fr->listedForcesGpu = listedForcesGpu.get();
1689 /* Initialize the mdAtoms structure.
1690 * mdAtoms is not filled with atom data,
1691 * as this can not be done now with domain decomposition.
1693 mdAtoms = makeMDAtoms(fplog, mtop, *inputrec, thisRankHasPmeGpuTask);
1694 if (globalState && thisRankHasPmeGpuTask)
1696 // The pinning of coordinates in the global state object works, because we only use
1697 // PME on GPU without DD or on a separate PME rank, and because the local state pointer
1698 // points to the global state object without DD.
1699 // FIXME: MD and EM separately set up the local state - this should happen in the same
1700 // function, which should also perform the pinning.
1701 changePinningPolicy(&globalState->x, pme_get_pinning_policy());
1704 /* Initialize the virtual site communication */
1705 vsite = makeVirtualSitesHandler(
1706 mtop, cr, fr->pbcType, updateGroups.updateGroupingPerMoleculeType());
1708 calc_shifts(box, fr->shift_vec);
1710 /* With periodic molecules the charge groups should be whole at start up
1711 * and the virtual sites should not be far from their proper positions.
1713 if (!inputrec->bContinuation && MASTER(cr)
1714 && !(inputrec->pbcType != PbcType::No && inputrec->bPeriodicMols))
1716 /* Make molecules whole at start of run */
1717 if (fr->pbcType != PbcType::No)
1719 do_pbc_first_mtop(fplog, inputrec->pbcType, box, &mtop, globalState->x.rvec_array());
1723 /* Correct initial vsite positions are required
1724 * for the initial distribution in the domain decomposition
1725 * and for the initial shell prediction.
1727 constructVirtualSitesGlobal(mtop, globalState->x);
1730 // Make the DD reverse topology, now that any vsites that are present are available
1731 if (DOMAINDECOMP(cr))
1733 dd_make_reverse_top(fplog, cr->dd, mtop, vsite.get(), *inputrec, domdecOptions.ddBondedChecking);
1736 if (EEL_PME(fr->ic->eeltype) || EVDW_PME(fr->ic->vdwtype))
1738 ewaldcoeff_q = fr->ic->ewaldcoeff_q;
1739 ewaldcoeff_lj = fr->ic->ewaldcoeff_lj;
1744 /* This is a PME only node */
1746 GMX_ASSERT(globalState == nullptr,
1747 "We don't need the state on a PME only rank and expect it to be unitialized");
1749 ewaldcoeff_q = calc_ewaldcoeff_q(inputrec->rcoulomb, inputrec->ewald_rtol);
1750 ewaldcoeff_lj = calc_ewaldcoeff_lj(inputrec->rvdw, inputrec->ewald_rtol_lj);
1753 gmx_pme_t* sepPmeData = nullptr;
1754 // This reference hides the fact that PME data is owned by runner on PME-only ranks and by forcerec on other ranks
1755 GMX_ASSERT(thisRankHasDuty(cr, DUTY_PP) == (fr != nullptr),
1756 "Double-checking that only PME-only ranks have no forcerec");
1757 gmx_pme_t*& pmedata = fr ? fr->pmedata : sepPmeData;
1759 // TODO should live in ewald module once its testing is improved
1761 // Later, this program could contain kernels that might be later
1762 // re-used as auto-tuning progresses, or subsequent simulations
1764 PmeGpuProgramStorage pmeGpuProgram;
1765 if (thisRankHasPmeGpuTask)
1768 (deviceStreamManager != nullptr),
1769 "GPU device stream manager should be initialized in order to use GPU for PME.");
1770 GMX_RELEASE_ASSERT((deviceInfo != nullptr),
1771 "GPU device should be initialized in order to use GPU for PME.");
1772 pmeGpuProgram = buildPmeGpuProgram(deviceStreamManager->context());
1775 /* Initiate PME if necessary,
1776 * either on all nodes or on dedicated PME nodes only. */
1777 if (EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype))
1779 if (mdAtoms && mdAtoms->mdatoms())
1781 nChargePerturbed = mdAtoms->mdatoms()->nChargePerturbed;
1782 if (EVDW_PME(inputrec->vdwtype))
1784 nTypePerturbed = mdAtoms->mdatoms()->nTypePerturbed;
1787 if (cr->npmenodes > 0)
1789 /* The PME only nodes need to know nChargePerturbed(FEP on Q) and nTypePerturbed(FEP on LJ)*/
1790 gmx_bcast(sizeof(nChargePerturbed), &nChargePerturbed, cr->mpi_comm_mysim);
1791 gmx_bcast(sizeof(nTypePerturbed), &nTypePerturbed, cr->mpi_comm_mysim);
1794 if (thisRankHasDuty(cr, DUTY_PME))
1798 // TODO: This should be in the builder.
1799 GMX_RELEASE_ASSERT(!runScheduleWork.simulationWork.useGpuPme
1800 || (deviceStreamManager != nullptr),
1801 "Device stream manager should be valid in order to use GPU "
1804 !runScheduleWork.simulationWork.useGpuPme
1805 || deviceStreamManager->streamIsValid(DeviceStreamType::Pme),
1806 "GPU PME stream should be valid in order to use GPU version of PME.");
1808 const DeviceContext* deviceContext = runScheduleWork.simulationWork.useGpuPme
1809 ? &deviceStreamManager->context()
1811 const DeviceStream* pmeStream =
1812 runScheduleWork.simulationWork.useGpuPme
1813 ? &deviceStreamManager->stream(DeviceStreamType::Pme)
1816 pmedata = gmx_pme_init(cr,
1817 getNumPmeDomains(cr->dd),
1819 nChargePerturbed != 0,
1820 nTypePerturbed != 0,
1821 mdrunOptions.reproducible,
1824 gmx_omp_nthreads_get(ModuleMultiThread::Pme),
1829 pmeGpuProgram.get(),
1832 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1837 if (EI_DYNAMICS(inputrec->eI))
1839 /* Turn on signal handling on all nodes */
1841 * (A user signal from the PME nodes (if any)
1842 * is communicated to the PP nodes.
1844 signal_handler_install();
1847 pull_t* pull_work = nullptr;
1848 if (thisRankHasDuty(cr, DUTY_PP))
1850 /* Assumes uniform use of the number of OpenMP threads */
1851 walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(ModuleMultiThread::Default));
1853 if (inputrec->bPull)
1855 /* Initialize pull code */
1856 pull_work = init_pull(fplog,
1857 inputrec->pull.get(),
1862 inputrec->fepvals->init_lambda);
1863 if (inputrec->pull->bXOutAverage || inputrec->pull->bFOutAverage)
1865 initPullHistory(pull_work, &observablesHistory);
1867 if (EI_DYNAMICS(inputrec->eI) && MASTER(cr))
1869 init_pull_output_files(pull_work, filenames.size(), filenames.data(), oenv, startingBehavior);
1873 std::unique_ptr<EnforcedRotation> enforcedRotation;
1876 /* Initialize enforced rotation code */
1877 enforcedRotation = init_rot(fplog,
1890 t_swap* swap = nullptr;
1891 if (inputrec->eSwapCoords != SwapType::No)
1893 /* Initialize ion swapping code */
1894 swap = init_swapcoords(fplog,
1896 opt2fn_master("-swap", filenames.size(), filenames.data(), cr),
1899 &observablesHistory,
1907 /* Let makeConstraints know whether we have essential dynamics constraints. */
1908 auto constr = makeConstraints(mtop,
1911 doEssentialDynamics,
1914 updateGroups.useUpdateGroups(),
1920 /* Energy terms and groups */
1921 gmx_enerdata_t enerd(mtop.groups.groups[SimulationAtomGroupType::EnergyOutput].size(),
1922 inputrec->fepvals->n_lambda);
1924 // cos acceleration is only supported by md, but older tpr
1925 // files might still combine it with other integrators
1926 GMX_RELEASE_ASSERT(inputrec->cos_accel == 0.0 || inputrec->eI == IntegrationAlgorithm::MD,
1927 "cos_acceleration is only supported by integrator=md");
1929 /* Kinetic energy data */
1930 gmx_ekindata_t ekind(inputrec->opts.ngtc,
1931 inputrec->cos_accel,
1932 gmx_omp_nthreads_get(ModuleMultiThread::Update));
1934 /* Set up interactive MD (IMD) */
1935 auto imdSession = makeImdSession(inputrec.get(),
1942 MASTER(cr) ? globalState->x : gmx::ArrayRef<gmx::RVec>(),
1946 mdrunOptions.imdOptions,
1949 if (DOMAINDECOMP(cr))
1951 GMX_RELEASE_ASSERT(fr, "fr was NULL while cr->duty was DUTY_PP");
1952 /* This call is not included in init_domain_decomposition
1953 * because fr->atomInfoForEachMoleculeBlock is set later.
1955 makeBondedLinks(cr->dd, mtop, fr->atomInfoForEachMoleculeBlock);
1958 if (runScheduleWork.simulationWork.useGpuBufferOps)
1960 fr->gpuForceReduction[gmx::AtomLocality::Local] = std::make_unique<gmx::GpuForceReduction>(
1961 deviceStreamManager->context(),
1962 deviceStreamManager->stream(gmx::DeviceStreamType::NonBondedLocal),
1964 fr->gpuForceReduction[gmx::AtomLocality::NonLocal] = std::make_unique<gmx::GpuForceReduction>(
1965 deviceStreamManager->context(),
1966 deviceStreamManager->stream(gmx::DeviceStreamType::NonBondedNonLocal),
1970 std::unique_ptr<gmx::StatePropagatorDataGpu> stateGpu;
1971 if (gpusWereDetected
1972 && ((runScheduleWork.simulationWork.useGpuPme && thisRankHasDuty(cr, DUTY_PME))
1973 || runScheduleWork.simulationWork.useGpuBufferOps))
1975 GpuApiCallBehavior transferKind =
1976 (inputrec->eI == IntegrationAlgorithm::MD && !doRerun && !useModularSimulator)
1977 ? GpuApiCallBehavior::Async
1978 : GpuApiCallBehavior::Sync;
1979 GMX_RELEASE_ASSERT(deviceStreamManager != nullptr,
1980 "GPU device stream manager should be initialized to use GPU.");
1981 stateGpu = std::make_unique<gmx::StatePropagatorDataGpu>(
1982 *deviceStreamManager, transferKind, pme_gpu_get_block_size(fr->pmedata), wcycle.get());
1983 fr->stateGpu = stateGpu.get();
1986 GMX_ASSERT(stopHandlerBuilder_, "Runner must provide StopHandlerBuilder to simulator.");
1987 SimulatorBuilder simulatorBuilder;
1989 simulatorBuilder.add(SimulatorStateData(globalState.get(), &observablesHistory, &enerd, &ekind));
1990 simulatorBuilder.add(std::move(membedHolder));
1991 simulatorBuilder.add(std::move(stopHandlerBuilder_));
1992 simulatorBuilder.add(SimulatorConfig(mdrunOptions, startingBehavior, &runScheduleWork));
1995 simulatorBuilder.add(SimulatorEnv(fplog, cr, ms, mdlog, oenv));
1996 simulatorBuilder.add(Profiling(&nrnb, walltime_accounting, wcycle.get()));
1997 simulatorBuilder.add(ConstraintsParam(
1998 constr.get(), enforcedRotation ? enforcedRotation->getLegacyEnfrot() : nullptr, vsite.get()));
1999 // TODO: Separate `fr` to a separate add, and make the `build` handle the coupling sensibly.
2000 simulatorBuilder.add(LegacyInput(
2001 static_cast<int>(filenames.size()), filenames.data(), inputrec.get(), fr.get()));
2002 simulatorBuilder.add(ReplicaExchangeParameters(replExParams));
2003 simulatorBuilder.add(InteractiveMD(imdSession.get()));
2004 simulatorBuilder.add(SimulatorModules(mdModules_->outputProvider(), mdModules_->notifiers()));
2005 simulatorBuilder.add(CenterOfMassPulling(pull_work));
2006 // Todo move to an MDModule
2007 simulatorBuilder.add(IonSwapping(swap));
2008 simulatorBuilder.add(TopologyData(mtop, mdAtoms.get()));
2009 simulatorBuilder.add(BoxDeformationHandle(deform.get()));
2010 simulatorBuilder.add(std::move(modularSimulatorCheckpointData));
2012 // build and run simulator object based on user-input
2013 auto simulator = simulatorBuilder.build(useModularSimulator);
2016 if (fr->pmePpCommGpu)
2018 // destroy object since it is no longer required. (This needs to be done while the GPU context still exists.)
2019 fr->pmePpCommGpu.reset();
2022 if (inputrec->bPull)
2024 finish_pull(pull_work);
2026 finish_swapcoords(swap);
2030 GMX_RELEASE_ASSERT(pmedata, "pmedata was NULL while cr->duty was not DUTY_PP");
2032 walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(ModuleMultiThread::Pme));
2033 gmx_pmeonly(pmedata,
2037 walltime_accounting,
2040 runScheduleWork.simulationWork.useGpuPmePpCommunication,
2041 deviceStreamManager.get());
2044 wallcycle_stop(wcycle.get(), WallCycleCounter::Run);
2046 /* Finish up, write some stuff
2047 * if rerunMD, don't write last frame again
2055 walltime_accounting,
2056 fr ? fr->nbv.get() : nullptr,
2058 EI_DYNAMICS(inputrec->eI) && !isMultiSim(ms));
2061 deviceStreamManager.reset(nullptr);
2065 gmx_pme_destroy(pmedata);
2069 // FIXME: this is only here to manually unpin mdAtoms->chargeA_ and state->x,
2070 // before we destroy the GPU context(s)
2071 // Pinned buffers are associated with contexts in CUDA.
2072 // As soon as we destroy GPU contexts after mdrunner() exits, these lines should go.
2073 mdAtoms.reset(nullptr);
2074 globalState.reset(nullptr);
2075 mdModules_.reset(nullptr); // destruct force providers here as they might also use the GPU
2076 listedForcesGpu.reset(nullptr);
2077 fr.reset(nullptr); // destruct forcerec before gpu
2078 // TODO convert to C++ so we can get rid of these frees
2081 if (!hwinfo_->deviceInfoList.empty())
2083 /* stop the GPU profiler (only CUDA) */
2087 /* With tMPI we need to wait for all ranks to finish deallocation before
2088 * destroying the CUDA context as some tMPI ranks may be sharing
2091 * This is not a concern in OpenCL where we use one context per rank.
2093 * Note: it is safe to not call the barrier on the ranks which do not use GPU,
2094 * but it is easier and more futureproof to call it on the whole node.
2096 * Note that this function needs to be called even if GPUs are not used
2097 * in this run because the PME ranks have no knowledge of whether GPUs
2098 * are used or not, but all ranks need to enter the barrier below.
2099 * \todo Remove this physical node barrier after making sure
2100 * that it's not needed anymore (with a shared GPU run).
2104 physicalNodeComm.barrier();
2107 if (!devFlags.usingCudaAwareMpi)
2109 // Don't reset GPU in case of CUDA-AWARE MPI
2110 // UCX creates CUDA buffers which are cleaned-up as part of MPI_Finalize()
2111 // resetting the device before MPI_Finalize() results in crashes inside UCX
2112 releaseDevice(deviceInfo);
2115 /* Does what it says */
2116 print_date_and_time(fplog, cr->nodeid, "Finished mdrun", gmx_gettime());
2117 walltime_accounting_destroy(walltime_accounting);
2119 // Ensure log file content is written
2122 gmx_fio_flush(logFileHandle);
2125 /* Reset FPEs (important for unit tests) by disabling them. Assumes no
2126 * exceptions were enabled before function was called. */
2129 gmx_fedisableexcept();
2132 auto rc = static_cast<int>(gmx_get_stop_condition());
2135 /* we need to join all threads. The sub-threads join when they
2136 exit this function, but the master thread needs to be told to
2146 Mdrunner::~Mdrunner()
2148 // Clean up of the Manager.
2149 // This will end up getting called on every thread-MPI rank, which is unnecessary,
2150 // but okay as long as threads synchronize some time before adding or accessing
2151 // a new set of restraints.
2152 if (restraintManager_)
2154 restraintManager_->clear();
2155 GMX_ASSERT(restraintManager_->countRestraints() == 0,
2156 "restraints added during runner life time should be cleared at runner "
2161 void Mdrunner::addPotential(std::shared_ptr<gmx::IRestraintPotential> puller, const std::string& name)
2163 GMX_ASSERT(restraintManager_, "Mdrunner must have a restraint manager.");
2164 // Not sure if this should be logged through the md logger or something else,
2165 // but it is helpful to have some sort of INFO level message sent somewhere.
2166 // std::cout << "Registering restraint named " << name << std::endl;
2168 // When multiple restraints are used, it may be wasteful to register them separately.
2169 // Maybe instead register an entire Restraint Manager as a force provider.
2170 restraintManager_->addToSpec(std::move(puller), name);
2173 Mdrunner::Mdrunner(std::unique_ptr<MDModules> mdModules) : mdModules_(std::move(mdModules)) {}
2175 Mdrunner::Mdrunner(Mdrunner&&) noexcept = default;
2177 //NOLINTNEXTLINE(performance-noexcept-move-constructor) working around GCC bug 58265 in CentOS 7
2178 Mdrunner& Mdrunner::operator=(Mdrunner&& /*handle*/) noexcept(BUGFREE_NOEXCEPT_STRING) = default;
2180 class Mdrunner::BuilderImplementation
2183 BuilderImplementation() = delete;
2184 BuilderImplementation(std::unique_ptr<MDModules> mdModules, compat::not_null<SimulationContext*> context);
2185 ~BuilderImplementation();
2187 BuilderImplementation& setExtraMdrunOptions(const MdrunOptions& options,
2188 real forceWarningThreshold,
2189 StartingBehavior startingBehavior);
2191 void addHardwareDetectionResult(const gmx_hw_info_t* hwinfo);
2193 void addDomdec(const DomdecOptions& options);
2195 void addInput(SimulationInputHandle inputHolder);
2197 void addVerletList(int nstlist);
2199 void addReplicaExchange(const ReplicaExchangeParameters& params);
2201 void addNonBonded(const char* nbpu_opt);
2203 void addPME(const char* pme_opt_, const char* pme_fft_opt_);
2205 void addBondedTaskAssignment(const char* bonded_opt);
2207 void addUpdateTaskAssignment(const char* update_opt);
2209 void addHardwareOptions(const gmx_hw_opt_t& hardwareOptions);
2211 void addFilenames(ArrayRef<const t_filenm> filenames);
2213 void addOutputEnvironment(gmx_output_env_t* outputEnvironment);
2215 void addLogFile(t_fileio* logFileHandle);
2217 void addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder);
2222 // Default parameters copied from runner.h
2223 // \todo Clarify source(s) of default parameters.
2225 const char* nbpu_opt_ = nullptr;
2226 const char* pme_opt_ = nullptr;
2227 const char* pme_fft_opt_ = nullptr;
2228 const char* bonded_opt_ = nullptr;
2229 const char* update_opt_ = nullptr;
2231 MdrunOptions mdrunOptions_;
2233 DomdecOptions domdecOptions_;
2235 ReplicaExchangeParameters replicaExchangeParameters_;
2237 //! Command-line override for the duration of a neighbor list with the Verlet scheme.
2240 //! World communicator, used for hardware detection and task assignment
2241 MPI_Comm libraryWorldCommunicator_ = MPI_COMM_NULL;
2243 //! Multisim communicator handle.
2244 gmx_multisim_t* multiSimulation_;
2246 //! mdrun communicator
2247 MPI_Comm simulationCommunicator_ = MPI_COMM_NULL;
2249 //! Print a warning if any force is larger than this (in kJ/mol nm).
2250 real forceWarningThreshold_ = -1;
2252 //! Whether the simulation will start afresh, or restart with/without appending.
2253 StartingBehavior startingBehavior_ = StartingBehavior::NewSimulation;
2255 //! The modules that comprise the functionality of mdrun.
2256 std::unique_ptr<MDModules> mdModules_;
2258 //! Detected hardware.
2259 const gmx_hw_info_t* hwinfo_ = nullptr;
2261 //! \brief Parallelism information.
2262 gmx_hw_opt_t hardwareOptions_;
2264 //! filename options for simulation.
2265 ArrayRef<const t_filenm> filenames_;
2267 /*! \brief Handle to output environment.
2269 * \todo gmx_output_env_t needs lifetime management.
2271 gmx_output_env_t* outputEnvironment_ = nullptr;
2273 /*! \brief Non-owning handle to MD log file.
2275 * \todo Context should own output facilities for client.
2276 * \todo Improve log file handle management.
2278 * Code managing the FILE* relies on the ability to set it to
2279 * nullptr to check whether the filehandle is valid.
2281 t_fileio* logFileHandle_ = nullptr;
2284 * \brief Builder for simulation stop signal handler.
2286 std::unique_ptr<StopHandlerBuilder> stopHandlerBuilder_ = nullptr;
2289 * \brief Sources for initial simulation state.
2291 * See issue #3652 for near-term refinements to the SimulationInput interface.
2293 * See issue #3379 for broader discussion on API aspects of simulation inputs and outputs.
2295 SimulationInputHandle inputHolder_;
2298 Mdrunner::BuilderImplementation::BuilderImplementation(std::unique_ptr<MDModules> mdModules,
2299 compat::not_null<SimulationContext*> context) :
2300 mdModules_(std::move(mdModules))
2302 libraryWorldCommunicator_ = context->libraryWorldCommunicator_;
2303 simulationCommunicator_ = context->simulationCommunicator_;
2304 multiSimulation_ = context->multiSimulation_.get();
2307 Mdrunner::BuilderImplementation::~BuilderImplementation() = default;
2309 Mdrunner::BuilderImplementation&
2310 Mdrunner::BuilderImplementation::setExtraMdrunOptions(const MdrunOptions& options,
2311 const real forceWarningThreshold,
2312 const StartingBehavior startingBehavior)
2314 mdrunOptions_ = options;
2315 forceWarningThreshold_ = forceWarningThreshold;
2316 startingBehavior_ = startingBehavior;
2320 void Mdrunner::BuilderImplementation::addDomdec(const DomdecOptions& options)
2322 domdecOptions_ = options;
2325 void Mdrunner::BuilderImplementation::addVerletList(int nstlist)
2330 void Mdrunner::BuilderImplementation::addReplicaExchange(const ReplicaExchangeParameters& params)
2332 replicaExchangeParameters_ = params;
2335 Mdrunner Mdrunner::BuilderImplementation::build()
2337 auto newRunner = Mdrunner(std::move(mdModules_));
2339 newRunner.mdrunOptions = mdrunOptions_;
2340 newRunner.pforce = forceWarningThreshold_;
2341 newRunner.startingBehavior = startingBehavior_;
2342 newRunner.domdecOptions = domdecOptions_;
2344 // \todo determine an invariant to check or confirm that all gmx_hw_opt_t objects are valid
2345 newRunner.hw_opt = hardwareOptions_;
2347 // No invariant to check. This parameter exists to optionally override other behavior.
2348 newRunner.nstlist_cmdline = nstlist_;
2350 newRunner.replExParams = replicaExchangeParameters_;
2352 newRunner.filenames = filenames_;
2354 newRunner.libraryWorldCommunicator = libraryWorldCommunicator_;
2356 newRunner.simulationCommunicator = simulationCommunicator_;
2358 // nullptr is a valid value for the multisim handle
2359 newRunner.ms = multiSimulation_;
2363 newRunner.hwinfo_ = hwinfo_;
2367 GMX_THROW(gmx::APIError(
2368 "MdrunnerBuilder::addHardwareDetectionResult() is required before build()"));
2373 newRunner.inputHolder_ = std::move(inputHolder_);
2377 GMX_THROW(gmx::APIError("MdrunnerBuilder::addInput() is required before build()."));
2380 // \todo Clarify ownership and lifetime management for gmx_output_env_t
2381 // \todo Update sanity checking when output environment has clearly specified invariants.
2382 // Initialization and default values for oenv are not well specified in the current version.
2383 if (outputEnvironment_)
2385 newRunner.oenv = outputEnvironment_;
2389 GMX_THROW(gmx::APIError(
2390 "MdrunnerBuilder::addOutputEnvironment() is required before build()"));
2393 newRunner.logFileHandle = logFileHandle_;
2397 newRunner.nbpu_opt = nbpu_opt_;
2401 GMX_THROW(gmx::APIError("MdrunnerBuilder::addNonBonded() is required before build()"));
2404 if (pme_opt_ && pme_fft_opt_)
2406 newRunner.pme_opt = pme_opt_;
2407 newRunner.pme_fft_opt = pme_fft_opt_;
2411 GMX_THROW(gmx::APIError("MdrunnerBuilder::addElectrostatics() is required before build()"));
2416 newRunner.bonded_opt = bonded_opt_;
2420 GMX_THROW(gmx::APIError(
2421 "MdrunnerBuilder::addBondedTaskAssignment() is required before build()"));
2426 newRunner.update_opt = update_opt_;
2430 GMX_THROW(gmx::APIError(
2431 "MdrunnerBuilder::addUpdateTaskAssignment() is required before build() "));
2435 newRunner.restraintManager_ = std::make_unique<gmx::RestraintManager>();
2437 if (stopHandlerBuilder_)
2439 newRunner.stopHandlerBuilder_ = std::move(stopHandlerBuilder_);
2443 newRunner.stopHandlerBuilder_ = std::make_unique<StopHandlerBuilder>();
2449 void Mdrunner::BuilderImplementation::addHardwareDetectionResult(const gmx_hw_info_t* hwinfo)
2454 void Mdrunner::BuilderImplementation::addNonBonded(const char* nbpu_opt)
2456 nbpu_opt_ = nbpu_opt;
2459 void Mdrunner::BuilderImplementation::addPME(const char* pme_opt, const char* pme_fft_opt)
2462 pme_fft_opt_ = pme_fft_opt;
2465 void Mdrunner::BuilderImplementation::addBondedTaskAssignment(const char* bonded_opt)
2467 bonded_opt_ = bonded_opt;
2470 void Mdrunner::BuilderImplementation::addUpdateTaskAssignment(const char* update_opt)
2472 update_opt_ = update_opt;
2475 void Mdrunner::BuilderImplementation::addHardwareOptions(const gmx_hw_opt_t& hardwareOptions)
2477 hardwareOptions_ = hardwareOptions;
2480 void Mdrunner::BuilderImplementation::addFilenames(ArrayRef<const t_filenm> filenames)
2482 filenames_ = filenames;
2485 void Mdrunner::BuilderImplementation::addOutputEnvironment(gmx_output_env_t* outputEnvironment)
2487 outputEnvironment_ = outputEnvironment;
2490 void Mdrunner::BuilderImplementation::addLogFile(t_fileio* logFileHandle)
2492 logFileHandle_ = logFileHandle;
2495 void Mdrunner::BuilderImplementation::addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder)
2497 stopHandlerBuilder_ = std::move(builder);
2500 void Mdrunner::BuilderImplementation::addInput(SimulationInputHandle inputHolder)
2502 inputHolder_ = std::move(inputHolder);
2505 MdrunnerBuilder::MdrunnerBuilder(std::unique_ptr<MDModules> mdModules,
2506 compat::not_null<SimulationContext*> context) :
2507 impl_{ std::make_unique<Mdrunner::BuilderImplementation>(std::move(mdModules), context) }
2511 MdrunnerBuilder::~MdrunnerBuilder() = default;
2513 MdrunnerBuilder& MdrunnerBuilder::addHardwareDetectionResult(const gmx_hw_info_t* hwinfo)
2515 impl_->addHardwareDetectionResult(hwinfo);
2519 MdrunnerBuilder& MdrunnerBuilder::addSimulationMethod(const MdrunOptions& options,
2520 real forceWarningThreshold,
2521 const StartingBehavior startingBehavior)
2523 impl_->setExtraMdrunOptions(options, forceWarningThreshold, startingBehavior);
2527 MdrunnerBuilder& MdrunnerBuilder::addDomainDecomposition(const DomdecOptions& options)
2529 impl_->addDomdec(options);
2533 MdrunnerBuilder& MdrunnerBuilder::addNeighborList(int nstlist)
2535 impl_->addVerletList(nstlist);
2539 MdrunnerBuilder& MdrunnerBuilder::addReplicaExchange(const ReplicaExchangeParameters& params)
2541 impl_->addReplicaExchange(params);
2545 MdrunnerBuilder& MdrunnerBuilder::addNonBonded(const char* nbpu_opt)
2547 impl_->addNonBonded(nbpu_opt);
2551 MdrunnerBuilder& MdrunnerBuilder::addElectrostatics(const char* pme_opt, const char* pme_fft_opt)
2553 // The builder method may become more general in the future, but in this version,
2554 // parameters for PME electrostatics are both required and the only parameters
2556 if (pme_opt && pme_fft_opt)
2558 impl_->addPME(pme_opt, pme_fft_opt);
2563 gmx::InvalidInputError("addElectrostatics() arguments must be non-null pointers."));
2568 MdrunnerBuilder& MdrunnerBuilder::addBondedTaskAssignment(const char* bonded_opt)
2570 impl_->addBondedTaskAssignment(bonded_opt);
2574 MdrunnerBuilder& MdrunnerBuilder::addUpdateTaskAssignment(const char* update_opt)
2576 impl_->addUpdateTaskAssignment(update_opt);
2580 Mdrunner MdrunnerBuilder::build()
2582 return impl_->build();
2585 MdrunnerBuilder& MdrunnerBuilder::addHardwareOptions(const gmx_hw_opt_t& hardwareOptions)
2587 impl_->addHardwareOptions(hardwareOptions);
2591 MdrunnerBuilder& MdrunnerBuilder::addFilenames(ArrayRef<const t_filenm> filenames)
2593 impl_->addFilenames(filenames);
2597 MdrunnerBuilder& MdrunnerBuilder::addOutputEnvironment(gmx_output_env_t* outputEnvironment)
2599 impl_->addOutputEnvironment(outputEnvironment);
2603 MdrunnerBuilder& MdrunnerBuilder::addLogFile(t_fileio* logFileHandle)
2605 impl_->addLogFile(logFileHandle);
2609 MdrunnerBuilder& MdrunnerBuilder::addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder)
2611 impl_->addStopHandlerBuilder(std::move(builder));
2615 MdrunnerBuilder& MdrunnerBuilder::addInput(SimulationInputHandle input)
2617 impl_->addInput(std::move(input));
2621 MdrunnerBuilder::MdrunnerBuilder(MdrunnerBuilder&&) noexcept = default;
2623 MdrunnerBuilder& MdrunnerBuilder::operator=(MdrunnerBuilder&&) noexcept = default;