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) 2011,2012,2013,2014,2015,2016,2017,2018,2019, 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
58 #include "gromacs/commandline/filenm.h"
59 #include "gromacs/compat/make_unique.h"
60 #include "gromacs/domdec/domdec.h"
61 #include "gromacs/domdec/domdec_struct.h"
62 #include "gromacs/domdec/localatomsetmanager.h"
63 #include "gromacs/ewald/ewald-utils.h"
64 #include "gromacs/ewald/pme.h"
65 #include "gromacs/ewald/pme-gpu-program.h"
66 #include "gromacs/fileio/checkpoint.h"
67 #include "gromacs/fileio/gmxfio.h"
68 #include "gromacs/fileio/oenv.h"
69 #include "gromacs/fileio/tpxio.h"
70 #include "gromacs/gmxlib/network.h"
71 #include "gromacs/gmxlib/nrnb.h"
72 #include "gromacs/gpu_utils/clfftinitializer.h"
73 #include "gromacs/gpu_utils/gpu_utils.h"
74 #include "gromacs/hardware/cpuinfo.h"
75 #include "gromacs/hardware/detecthardware.h"
76 #include "gromacs/hardware/printhardware.h"
77 #include "gromacs/listed-forces/disre.h"
78 #include "gromacs/listed-forces/gpubonded.h"
79 #include "gromacs/listed-forces/orires.h"
80 #include "gromacs/math/functions.h"
81 #include "gromacs/math/utilities.h"
82 #include "gromacs/math/vec.h"
83 #include "gromacs/mdlib/boxdeformation.h"
84 #include "gromacs/mdlib/calc_verletbuf.h"
85 #include "gromacs/mdlib/forcerec.h"
86 #include "gromacs/mdlib/gmx_omp_nthreads.h"
87 #include "gromacs/mdlib/makeconstraints.h"
88 #include "gromacs/mdlib/md_support.h"
89 #include "gromacs/mdlib/mdatoms.h"
90 #include "gromacs/mdlib/mdrun.h"
91 #include "gromacs/mdlib/membed.h"
92 #include "gromacs/mdlib/nb_verlet.h"
93 #include "gromacs/mdlib/nbnxn_gpu_data_mgmt.h"
94 #include "gromacs/mdlib/nbnxn_search.h"
95 #include "gromacs/mdlib/nbnxn_tuning.h"
96 #include "gromacs/mdlib/ppforceworkload.h"
97 #include "gromacs/mdlib/qmmm.h"
98 #include "gromacs/mdlib/sighandler.h"
99 #include "gromacs/mdlib/sim_util.h"
100 #include "gromacs/mdlib/stophandler.h"
101 #include "gromacs/mdrun/legacymdrunoptions.h"
102 #include "gromacs/mdrun/logging.h"
103 #include "gromacs/mdrun/multisim.h"
104 #include "gromacs/mdrun/simulationcontext.h"
105 #include "gromacs/mdrunutility/mdmodules.h"
106 #include "gromacs/mdrunutility/threadaffinity.h"
107 #include "gromacs/mdtypes/commrec.h"
108 #include "gromacs/mdtypes/fcdata.h"
109 #include "gromacs/mdtypes/inputrec.h"
110 #include "gromacs/mdtypes/md_enums.h"
111 #include "gromacs/mdtypes/observableshistory.h"
112 #include "gromacs/mdtypes/state.h"
113 #include "gromacs/pbcutil/pbc.h"
114 #include "gromacs/pulling/output.h"
115 #include "gromacs/pulling/pull.h"
116 #include "gromacs/pulling/pull_rotation.h"
117 #include "gromacs/restraint/manager.h"
118 #include "gromacs/restraint/restraintmdmodule.h"
119 #include "gromacs/restraint/restraintpotential.h"
120 #include "gromacs/swap/swapcoords.h"
121 #include "gromacs/taskassignment/decidegpuusage.h"
122 #include "gromacs/taskassignment/resourcedivision.h"
123 #include "gromacs/taskassignment/taskassignment.h"
124 #include "gromacs/taskassignment/usergpuids.h"
125 #include "gromacs/timing/wallcycle.h"
126 #include "gromacs/topology/mtop_util.h"
127 #include "gromacs/trajectory/trajectoryframe.h"
128 #include "gromacs/utility/basenetwork.h"
129 #include "gromacs/utility/cstringutil.h"
130 #include "gromacs/utility/exceptions.h"
131 #include "gromacs/utility/fatalerror.h"
132 #include "gromacs/utility/filestream.h"
133 #include "gromacs/utility/gmxassert.h"
134 #include "gromacs/utility/gmxmpi.h"
135 #include "gromacs/utility/logger.h"
136 #include "gromacs/utility/loggerbuilder.h"
137 #include "gromacs/utility/physicalnodecommunicator.h"
138 #include "gromacs/utility/pleasecite.h"
139 #include "gromacs/utility/programcontext.h"
140 #include "gromacs/utility/smalloc.h"
141 #include "gromacs/utility/stringutil.h"
143 #include "integrator.h"
144 #include "replicaexchange.h"
147 #include "corewrap.h"
153 /*! \brief Barrier for safe simultaneous thread access to mdrunner data
155 * Used to ensure that the master thread does not modify mdrunner during copy
156 * on the spawned threads. */
157 static void threadMpiMdrunnerAccessBarrier()
160 MPI_Barrier(MPI_COMM_WORLD);
164 Mdrunner Mdrunner::cloneOnSpawnedThread() const
166 auto newRunner = Mdrunner();
168 // All runners in the same process share a restraint manager resource because it is
169 // part of the interface to the client code, which is associated only with the
170 // original thread. Handles to the same resources can be obtained by copy.
172 newRunner.restraintManager_ = compat::make_unique<RestraintManager>(*restraintManager_);
175 // Copy original cr pointer before master thread can pass the thread barrier
176 newRunner.cr = reinitialize_commrec_for_this_thread(cr);
178 // Copy members of master runner.
179 // \todo Replace with builder when Simulation context and/or runner phases are better defined.
180 // Ref https://redmine.gromacs.org/issues/2587 and https://redmine.gromacs.org/issues/2375
181 newRunner.hw_opt = hw_opt;
182 newRunner.filenames = filenames;
184 newRunner.oenv = oenv;
185 newRunner.mdrunOptions = mdrunOptions;
186 newRunner.domdecOptions = domdecOptions;
187 newRunner.nbpu_opt = nbpu_opt;
188 newRunner.pme_opt = pme_opt;
189 newRunner.pme_fft_opt = pme_fft_opt;
190 newRunner.bonded_opt = bonded_opt;
191 newRunner.nstlist_cmdline = nstlist_cmdline;
192 newRunner.replExParams = replExParams;
193 newRunner.pforce = pforce;
195 newRunner.stopHandlerBuilder_ = compat::make_unique<StopHandlerBuilder>(*stopHandlerBuilder_);
197 threadMpiMdrunnerAccessBarrier();
199 GMX_RELEASE_ASSERT(!MASTER(newRunner.cr), "cloneOnSpawnedThread should only be called on spawned threads");
204 /*! \brief The callback used for running on spawned threads.
206 * Obtains the pointer to the master mdrunner object from the one
207 * argument permitted to the thread-launch API call, copies it to make
208 * a new runner for this thread, reinitializes necessary data, and
209 * proceeds to the simulation. */
210 static void mdrunner_start_fn(const void *arg)
214 auto masterMdrunner = reinterpret_cast<const gmx::Mdrunner *>(arg);
215 /* copy the arg list to make sure that it's thread-local. This
216 doesn't copy pointed-to items, of course; fnm, cr and fplog
217 are reset in the call below, all others should be const. */
218 gmx::Mdrunner mdrunner = masterMdrunner->cloneOnSpawnedThread();
221 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
225 /*! \brief Start thread-MPI threads.
227 * Called by mdrunner() to start a specific number of threads
228 * (including the main thread) for thread-parallel runs. This in turn
229 * calls mdrunner() for each thread. All options are the same as for
231 t_commrec *Mdrunner::spawnThreads(int numThreadsToLaunch) const
234 /* first check whether we even need to start tMPI */
235 if (numThreadsToLaunch < 2)
241 /* now spawn new threads that start mdrunner_start_fn(), while
242 the main thread returns, we set thread affinity later */
243 if (tMPI_Init_fn(TRUE, numThreadsToLaunch, TMPI_AFFINITY_NONE,
244 mdrunner_start_fn, static_cast<const void*>(this)) != TMPI_SUCCESS)
246 GMX_THROW(gmx::InternalError("Failed to spawn thread-MPI threads"));
249 threadMpiMdrunnerAccessBarrier();
251 GMX_UNUSED_VALUE(mdrunner_start_fn);
254 return reinitialize_commrec_for_this_thread(cr);
259 /*! \brief Initialize variables for Verlet scheme simulation */
260 static void prepare_verlet_scheme(FILE *fplog,
264 const gmx_mtop_t *mtop,
266 bool makeGpuPairList,
267 const gmx::CpuInfo &cpuinfo)
269 /* For NVE simulations, we will retain the initial list buffer */
270 if (EI_DYNAMICS(ir->eI) &&
271 ir->verletbuf_tol > 0 &&
272 !(EI_MD(ir->eI) && ir->etc == etcNO))
274 /* Update the Verlet buffer size for the current run setup */
276 /* Here we assume SIMD-enabled kernels are being used. But as currently
277 * calc_verlet_buffer_size gives the same results for 4x8 and 4x4
278 * and 4x2 gives a larger buffer than 4x4, this is ok.
280 ListSetupType listType = (makeGpuPairList ? ListSetupType::Gpu : ListSetupType::CpuSimdWhenSupported);
281 VerletbufListSetup listSetup = verletbufGetSafeListSetup(listType);
284 calc_verlet_buffer_size(mtop, det(box), ir, ir->nstlist, ir->nstlist - 1, -1, &listSetup, nullptr, &rlist_new);
286 if (rlist_new != ir->rlist)
288 if (fplog != nullptr)
290 fprintf(fplog, "\nChanging rlist from %g to %g for non-bonded %dx%d atom kernels\n\n",
291 ir->rlist, rlist_new,
292 listSetup.cluster_size_i, listSetup.cluster_size_j);
294 ir->rlist = rlist_new;
298 if (nstlist_cmdline > 0 && (!EI_DYNAMICS(ir->eI) || ir->verletbuf_tol <= 0))
300 gmx_fatal(FARGS, "Can not set nstlist without %s",
301 !EI_DYNAMICS(ir->eI) ? "dynamics" : "verlet-buffer-tolerance");
304 if (EI_DYNAMICS(ir->eI))
306 /* Set or try nstlist values */
307 increaseNstlist(fplog, cr, ir, nstlist_cmdline, mtop, box, makeGpuPairList, cpuinfo);
311 /*! \brief Override the nslist value in inputrec
313 * with value passed on the command line (if any)
315 static void override_nsteps_cmdline(const gmx::MDLogger &mdlog,
316 int64_t nsteps_cmdline,
321 /* override with anything else than the default -2 */
322 if (nsteps_cmdline > -2)
324 char sbuf_steps[STEPSTRSIZE];
325 char sbuf_msg[STRLEN];
327 ir->nsteps = nsteps_cmdline;
328 if (EI_DYNAMICS(ir->eI) && nsteps_cmdline != -1)
330 sprintf(sbuf_msg, "Overriding nsteps with value passed on the command line: %s steps, %.3g ps",
331 gmx_step_str(nsteps_cmdline, sbuf_steps),
332 fabs(nsteps_cmdline*ir->delta_t));
336 sprintf(sbuf_msg, "Overriding nsteps with value passed on the command line: %s steps",
337 gmx_step_str(nsteps_cmdline, sbuf_steps));
340 GMX_LOG(mdlog.warning).asParagraph().appendText(sbuf_msg);
342 else if (nsteps_cmdline < -2)
344 gmx_fatal(FARGS, "Invalid nsteps value passed on the command line: %" PRId64,
347 /* Do nothing if nsteps_cmdline == -2 */
353 /*! \brief Return whether GPU acceleration of nonbondeds is supported with the given settings.
355 * If not, and if a warning may be issued, logs a warning about
356 * falling back to CPU code. With thread-MPI, only the first
357 * call to this function should have \c issueWarning true. */
358 static bool gpuAccelerationOfNonbondedIsUseful(const MDLogger &mdlog,
359 const t_inputrec *ir,
362 if (ir->opts.ngener - ir->nwall > 1)
364 /* The GPU code does not support more than one energy group.
365 * If the user requested GPUs explicitly, a fatal error is given later.
369 GMX_LOG(mdlog.warning).asParagraph()
370 .appendText("Multiple energy groups is not implemented for GPUs, falling back to the CPU. "
371 "For better performance, run on the GPU without energy groups and then do "
372 "gmx mdrun -rerun option on the trajectory with an energy group .tpr file.");
379 //! Initializes the logger for mdrun.
380 static gmx::LoggerOwner buildLogger(FILE *fplog, const t_commrec *cr)
382 gmx::LoggerBuilder builder;
383 if (fplog != nullptr)
385 builder.addTargetFile(gmx::MDLogger::LogLevel::Info, fplog);
387 if (cr == nullptr || SIMMASTER(cr))
389 builder.addTargetStream(gmx::MDLogger::LogLevel::Warning,
390 &gmx::TextOutputFile::standardError());
392 return builder.build();
395 //! Make a TaskTarget from an mdrun argument string.
396 static TaskTarget findTaskTarget(const char *optionString)
398 TaskTarget returnValue = TaskTarget::Auto;
400 if (strncmp(optionString, "auto", 3) == 0)
402 returnValue = TaskTarget::Auto;
404 else if (strncmp(optionString, "cpu", 3) == 0)
406 returnValue = TaskTarget::Cpu;
408 else if (strncmp(optionString, "gpu", 3) == 0)
410 returnValue = TaskTarget::Gpu;
414 GMX_ASSERT(false, "Option string should have been checked for sanity already");
420 int Mdrunner::mdrunner()
424 t_forcerec *fr = nullptr;
425 t_fcdata *fcd = nullptr;
426 real ewaldcoeff_q = 0;
427 real ewaldcoeff_lj = 0;
428 int nChargePerturbed = -1, nTypePerturbed = 0;
429 gmx_wallcycle_t wcycle;
430 gmx_walltime_accounting_t walltime_accounting = nullptr;
432 int64_t reset_counters;
433 int nthreads_pme = 1;
434 gmx_membed_t * membed = nullptr;
435 gmx_hw_info_t *hwinfo = nullptr;
437 /* CAUTION: threads may be started later on in this function, so
438 cr doesn't reflect the final parallel state right now */
439 std::unique_ptr<gmx::MDModules> mdModules(new gmx::MDModules);
440 t_inputrec inputrecInstance;
441 t_inputrec *inputrec = &inputrecInstance;
444 bool doMembed = opt2bSet("-membed", filenames.size(), filenames.data());
445 bool doRerun = mdrunOptions.rerun;
447 // Handle task-assignment related user options.
448 EmulateGpuNonbonded emulateGpuNonbonded = (getenv("GMX_EMULATE_GPU") != nullptr ?
449 EmulateGpuNonbonded::Yes : EmulateGpuNonbonded::No);
450 std::vector<int> gpuIdsAvailable;
453 gpuIdsAvailable = parseUserGpuIds(hw_opt.gpuIdsAvailable);
454 // TODO We could put the GPU IDs into a std::map to find
455 // duplicates, but for the small numbers of IDs involved, this
456 // code is simple and fast.
457 for (size_t i = 0; i != gpuIdsAvailable.size(); ++i)
459 for (size_t j = i+1; j != gpuIdsAvailable.size(); ++j)
461 if (gpuIdsAvailable[i] == gpuIdsAvailable[j])
463 GMX_THROW(InvalidInputError(formatString("The string of available GPU device IDs '%s' may not contain duplicate device IDs", hw_opt.gpuIdsAvailable.c_str())));
468 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
470 std::vector<int> userGpuTaskAssignment;
473 userGpuTaskAssignment = parseUserGpuIds(hw_opt.userGpuTaskAssignment);
475 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
476 auto nonbondedTarget = findTaskTarget(nbpu_opt);
477 auto pmeTarget = findTaskTarget(pme_opt);
478 auto pmeFftTarget = findTaskTarget(pme_fft_opt);
479 auto bondedTarget = findTaskTarget(bonded_opt);
480 PmeRunMode pmeRunMode = PmeRunMode::None;
482 // Here we assume that SIMMASTER(cr) does not change even after the
483 // threads are started.
485 FILE *fplog = nullptr;
486 // If we are appending, we don't write log output because we need
487 // to check that the old log file matches what the checkpoint file
488 // expects. Otherwise, we should start to write log output now if
489 // there is a file ready for it.
490 if (logFileHandle != nullptr && !mdrunOptions.continuationOptions.appendFiles)
492 fplog = gmx_fio_getfp(logFileHandle);
494 gmx::LoggerOwner logOwner(buildLogger(fplog, cr));
495 gmx::MDLogger mdlog(logOwner.logger());
497 // TODO The thread-MPI master rank makes a working
498 // PhysicalNodeCommunicator here, but it gets rebuilt by all ranks
499 // after the threads have been launched. This works because no use
500 // is made of that communicator until after the execution paths
501 // have rejoined. But it is likely that we can improve the way
502 // this is expressed, e.g. by expressly running detection only the
503 // master rank for thread-MPI, rather than relying on the mutex
504 // and reference count.
505 PhysicalNodeCommunicator physicalNodeComm(MPI_COMM_WORLD, gmx_physicalnode_id_hash());
506 hwinfo = gmx_detect_hardware(mdlog, physicalNodeComm);
508 gmx_print_detected_hardware(fplog, cr, ms, mdlog, hwinfo);
510 std::vector<int> gpuIdsToUse;
511 auto compatibleGpus = getCompatibleGpus(hwinfo->gpu_info);
512 if (gpuIdsAvailable.empty())
514 gpuIdsToUse = compatibleGpus;
518 for (const auto &availableGpuId : gpuIdsAvailable)
520 bool availableGpuIsCompatible = false;
521 for (const auto &compatibleGpuId : compatibleGpus)
523 if (availableGpuId == compatibleGpuId)
525 availableGpuIsCompatible = true;
529 if (!availableGpuIsCompatible)
531 gmx_fatal(FARGS, "You limited the set of compatible GPUs to a set that included ID #%d, but that ID is not for a compatible GPU. List only compatible GPUs.", availableGpuId);
533 gpuIdsToUse.push_back(availableGpuId);
537 if (fplog != nullptr)
539 /* Print references after all software/hardware printing */
540 please_cite(fplog, "Abraham2015");
541 please_cite(fplog, "Pall2015");
542 please_cite(fplog, "Pronk2013");
543 please_cite(fplog, "Hess2008b");
544 please_cite(fplog, "Spoel2005a");
545 please_cite(fplog, "Lindahl2001a");
546 please_cite(fplog, "Berendsen95a");
547 writeSourceDoi(fplog);
550 std::unique_ptr<t_state> globalState;
554 /* Only the master rank has the global state */
555 globalState = compat::make_unique<t_state>();
557 /* Read (nearly) all data required for the simulation */
558 read_tpx_state(ftp2fn(efTPR, filenames.size(), filenames.data()), inputrec, globalState.get(), &mtop);
560 /* In rerun, set velocities to zero if present */
561 if (doRerun && ((globalState->flags & (1 << estV)) != 0))
563 // rerun does not use velocities
564 GMX_LOG(mdlog.info).asParagraph().appendText(
565 "Rerun trajectory contains velocities. Rerun does only evaluate "
566 "potential energy and forces. The velocities will be ignored.");
567 for (int i = 0; i < globalState->natoms; i++)
569 clear_rvec(globalState->v[i]);
571 globalState->flags &= ~(1 << estV);
574 if (inputrec->cutoff_scheme != ecutsVERLET)
576 if (nstlist_cmdline > 0)
578 gmx_fatal(FARGS, "Can not set nstlist with the group cut-off scheme");
581 if (!compatibleGpus.empty())
583 GMX_LOG(mdlog.warning).asParagraph().appendText(
584 "NOTE: GPU(s) found, but the current simulation can not use GPUs\n"
585 " To use a GPU, set the mdp option: cutoff-scheme = Verlet");
590 /* Check and update the hardware options for internal consistency */
591 check_and_update_hw_opt_1(mdlog, &hw_opt, cr, domdecOptions.numPmeRanks);
593 /* Early check for externally set process affinity. */
594 gmx_check_thread_affinity_set(mdlog, cr,
595 &hw_opt, hwinfo->nthreads_hw_avail, FALSE);
597 if (GMX_THREAD_MPI && SIMMASTER(cr))
599 if (domdecOptions.numPmeRanks > 0 && hw_opt.nthreads_tmpi <= 0)
601 gmx_fatal(FARGS, "You need to explicitly specify the number of MPI threads (-ntmpi) when using separate PME ranks");
604 /* Since the master knows the cut-off scheme, update hw_opt for this.
605 * This is done later for normal MPI and also once more with tMPI
606 * for all tMPI ranks.
608 check_and_update_hw_opt_2(&hw_opt, inputrec->cutoff_scheme);
610 bool useGpuForNonbonded = false;
611 bool useGpuForPme = false;
614 // If the user specified the number of ranks, then we must
615 // respect that, but in default mode, we need to allow for
616 // the number of GPUs to choose the number of ranks.
617 auto canUseGpuForNonbonded = buildSupportsNonbondedOnGpu(nullptr);
618 useGpuForNonbonded = decideWhetherToUseGpusForNonbondedWithThreadMpi
619 (nonbondedTarget, gpuIdsToUse, userGpuTaskAssignment, emulateGpuNonbonded,
620 canUseGpuForNonbonded,
621 inputrec->cutoff_scheme == ecutsVERLET,
622 gpuAccelerationOfNonbondedIsUseful(mdlog, inputrec, GMX_THREAD_MPI),
623 hw_opt.nthreads_tmpi);
624 auto canUseGpuForPme = pme_gpu_supports_build(*hwinfo, nullptr) && pme_gpu_supports_input(*inputrec, mtop, nullptr);
625 useGpuForPme = decideWhetherToUseGpusForPmeWithThreadMpi
626 (useGpuForNonbonded, pmeTarget, gpuIdsToUse, userGpuTaskAssignment,
627 canUseGpuForPme, hw_opt.nthreads_tmpi, domdecOptions.numPmeRanks);
630 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
632 /* Determine how many thread-MPI ranks to start.
634 * TODO Over-writing the user-supplied value here does
635 * prevent any possible subsequent checks from working
637 hw_opt.nthreads_tmpi = get_nthreads_mpi(hwinfo,
646 // Now start the threads for thread MPI.
647 cr = spawnThreads(hw_opt.nthreads_tmpi);
648 /* The main thread continues here with a new cr. We don't deallocate
649 the old cr because other threads may still be reading it. */
650 // TODO Both master and spawned threads call dup_tfn and
651 // reinitialize_commrec_for_this_thread. Find a way to express
653 physicalNodeComm = PhysicalNodeCommunicator(MPI_COMM_WORLD, gmx_physicalnode_id_hash());
655 // END OF CAUTION: cr and physicalNodeComm are now reliable
659 /* now broadcast everything to the non-master nodes/threads: */
660 init_parallel(cr, inputrec, &mtop);
663 // Now each rank knows the inputrec that SIMMASTER read and used,
664 // and (if applicable) cr->nnodes has been assigned the number of
665 // thread-MPI ranks that have been chosen. The ranks can now all
666 // run the task-deciding functions and will agree on the result
667 // without needing to communicate.
669 // TODO Should we do the communication in debug mode to support
670 // having an assertion?
672 // Note that these variables describe only their own node.
674 // Note that when bonded interactions run on a GPU they always run
675 // alongside a nonbonded task, so do not influence task assignment
676 // even though they affect the force calculation workload.
677 bool useGpuForNonbonded = false;
678 bool useGpuForPme = false;
679 bool useGpuForBonded = false;
682 // It's possible that there are different numbers of GPUs on
683 // different nodes, which is the user's responsibilty to
684 // handle. If unsuitable, we will notice that during task
686 bool gpusWereDetected = hwinfo->ngpu_compatible_tot > 0;
687 bool usingVerletScheme = inputrec->cutoff_scheme == ecutsVERLET;
688 auto canUseGpuForNonbonded = buildSupportsNonbondedOnGpu(nullptr);
689 useGpuForNonbonded = decideWhetherToUseGpusForNonbonded(nonbondedTarget, userGpuTaskAssignment,
691 canUseGpuForNonbonded,
693 gpuAccelerationOfNonbondedIsUseful(mdlog, inputrec, !GMX_THREAD_MPI),
695 auto canUseGpuForPme = pme_gpu_supports_build(*hwinfo, nullptr) && pme_gpu_supports_input(*inputrec, mtop, nullptr);
696 useGpuForPme = decideWhetherToUseGpusForPme(useGpuForNonbonded, pmeTarget, userGpuTaskAssignment,
697 canUseGpuForPme, cr->nnodes, domdecOptions.numPmeRanks,
699 auto canUseGpuForBonded = buildSupportsGpuBondeds(nullptr) && inputSupportsGpuBondeds(*inputrec, mtop, nullptr);
701 decideWhetherToUseGpusForBonded(useGpuForNonbonded, useGpuForPme, usingVerletScheme,
702 bondedTarget, canUseGpuForBonded,
703 EVDW_PME(inputrec->vdwtype),
704 EEL_PME_EWALD(inputrec->coulombtype),
705 domdecOptions.numPmeRanks, gpusWereDetected);
707 pmeRunMode = (useGpuForPme ? PmeRunMode::GPU : PmeRunMode::CPU);
708 if (pmeRunMode == PmeRunMode::GPU)
710 if (pmeFftTarget == TaskTarget::Cpu)
712 pmeRunMode = PmeRunMode::Mixed;
715 else if (pmeFftTarget == TaskTarget::Gpu)
717 gmx_fatal(FARGS, "Assigning FFTs to GPU requires PME to be assigned to GPU as well. With PME on CPU you should not be using -pmefft.");
720 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
723 // TODO: hide restraint implementation details from Mdrunner.
724 // There is nothing unique about restraints at this point as far as the
725 // Mdrunner is concerned. The Mdrunner should just be getting a sequence of
726 // factory functions from the SimulationContext on which to call mdModules->add().
727 // TODO: capture all restraints into a single RestraintModule, passed to the runner builder.
728 for (auto && restraint : restraintManager_->getRestraints())
730 auto module = RestraintMDModule::create(restraint,
732 mdModules->add(std::move(module));
735 // TODO: Error handling
736 mdModules->assignOptionsToModules(*inputrec->params, nullptr);
738 if (fplog != nullptr)
740 pr_inputrec(fplog, 0, "Input Parameters", inputrec, FALSE);
741 fprintf(fplog, "\n");
746 /* now make sure the state is initialized and propagated */
747 set_state_entries(globalState.get(), inputrec);
750 /* NM and TPI parallelize over force/energy calculations, not atoms,
751 * so we need to initialize and broadcast the global state.
753 if (inputrec->eI == eiNM || inputrec->eI == eiTPI)
757 globalState = compat::make_unique<t_state>();
759 broadcastStateWithoutDynamics(cr, globalState.get());
762 /* A parallel command line option consistency check that we can
763 only do after any threads have started. */
764 if (!PAR(cr) && (domdecOptions.numCells[XX] > 1 ||
765 domdecOptions.numCells[YY] > 1 ||
766 domdecOptions.numCells[ZZ] > 1 ||
767 domdecOptions.numPmeRanks > 0))
770 "The -dd or -npme option request a parallel simulation, "
772 "but %s was compiled without threads or MPI enabled", output_env_get_program_display_name(oenv));
775 "but the number of MPI-threads (option -ntmpi) is not set or is 1");
777 "but %s was not started through mpirun/mpiexec or only one rank was requested through mpirun/mpiexec", output_env_get_program_display_name(oenv));
783 (EI_ENERGY_MINIMIZATION(inputrec->eI) || eiNM == inputrec->eI))
785 gmx_fatal(FARGS, "The .mdp file specified an energy mininization or normal mode algorithm, and these are not compatible with mdrun -rerun");
788 if (can_use_allvsall(inputrec, TRUE, cr, fplog) && DOMAINDECOMP(cr))
790 gmx_fatal(FARGS, "All-vs-all loops do not work with domain decomposition, use a single MPI rank");
793 if (!(EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype)))
795 if (domdecOptions.numPmeRanks > 0)
797 gmx_fatal_collective(FARGS, cr->mpi_comm_mysim, MASTER(cr),
798 "PME-only ranks are requested, but the system does not use PME for electrostatics or LJ");
801 domdecOptions.numPmeRanks = 0;
804 if (useGpuForNonbonded && domdecOptions.numPmeRanks < 0)
806 /* With NB GPUs we don't automatically use PME-only CPU ranks. PME ranks can
807 * improve performance with many threads per GPU, since our OpenMP
808 * scaling is bad, but it's difficult to automate the setup.
810 domdecOptions.numPmeRanks = 0;
814 if (domdecOptions.numPmeRanks < 0)
816 domdecOptions.numPmeRanks = 0;
817 // TODO possibly print a note that one can opt-in for a separate PME GPU rank?
821 GMX_RELEASE_ASSERT(domdecOptions.numPmeRanks <= 1, "PME GPU decomposition is not supported");
828 fcRegisterSteps(inputrec->nsteps, inputrec->init_step);
832 /* NMR restraints must be initialized before load_checkpoint,
833 * since with time averaging the history is added to t_state.
834 * For proper consistency check we therefore need to extend
836 * So the PME-only nodes (if present) will also initialize
837 * the distance restraints.
841 /* This needs to be called before read_checkpoint to extend the state */
842 init_disres(fplog, &mtop, inputrec, cr, ms, fcd, globalState.get(), replExParams.exchangeInterval > 0);
844 init_orires(fplog, &mtop, inputrec, cr, ms, globalState.get(), &(fcd->orires));
846 auto deform = prepareBoxDeformation(globalState->box, cr, *inputrec);
848 ObservablesHistory observablesHistory = {};
850 ContinuationOptions &continuationOptions = mdrunOptions.continuationOptions;
852 if (continuationOptions.startedFromCheckpoint)
854 /* Check if checkpoint file exists before doing continuation.
855 * This way we can use identical input options for the first and subsequent runs...
859 load_checkpoint(opt2fn_master("-cpi", filenames.size(), filenames.data(), cr),
861 cr, domdecOptions.numCells,
862 inputrec, globalState.get(),
863 &bReadEkin, &observablesHistory,
864 continuationOptions.appendFiles,
865 continuationOptions.appendFilesOptionSet,
866 mdrunOptions.reproducible);
870 continuationOptions.haveReadEkin = true;
873 if (continuationOptions.appendFiles && logFileHandle)
875 // Now we can start normal logging to the truncated log file.
876 fplog = gmx_fio_getfp(logFileHandle);
877 prepareLogAppending(fplog);
878 logOwner = buildLogger(fplog, cr);
879 mdlog = logOwner.logger();
883 if (mdrunOptions.numStepsCommandline > -2)
885 GMX_LOG(mdlog.info).asParagraph().
886 appendText("The -nsteps functionality is deprecated, and may be removed in a future version. "
887 "Consider using gmx convert-tpr -nsteps or changing the appropriate .mdp file field.");
889 /* override nsteps with value set on the commamdline */
890 override_nsteps_cmdline(mdlog, mdrunOptions.numStepsCommandline, inputrec);
894 copy_mat(globalState->box, box);
899 gmx_bcast(sizeof(box), box, cr);
902 /* Update rlist and nstlist. */
903 if (inputrec->cutoff_scheme == ecutsVERLET)
905 prepare_verlet_scheme(fplog, cr, inputrec, nstlist_cmdline, &mtop, box,
906 useGpuForNonbonded || (emulateGpuNonbonded == EmulateGpuNonbonded::Yes), *hwinfo->cpuInfo);
909 LocalAtomSetManager atomSets;
911 if (PAR(cr) && !(EI_TPI(inputrec->eI) ||
912 inputrec->eI == eiNM))
914 cr->dd = init_domain_decomposition(mdlog, cr, domdecOptions, mdrunOptions,
916 box, positionsFromStatePointer(globalState.get()),
918 // Note that local state still does not exist yet.
922 /* PME, if used, is done on all nodes with 1D decomposition */
924 cr->duty = (DUTY_PP | DUTY_PME);
926 if (inputrec->ePBC == epbcSCREW)
929 "pbc=screw is only implemented with domain decomposition");
935 /* After possible communicator splitting in make_dd_communicators.
936 * we can set up the intra/inter node communication.
938 gmx_setup_nodecomm(fplog, cr);
944 GMX_LOG(mdlog.warning).asParagraph().appendTextFormatted(
945 "This is simulation %d out of %d running as a composite GROMACS\n"
946 "multi-simulation job. Setup for this simulation:\n",
949 GMX_LOG(mdlog.warning).appendTextFormatted(
953 cr->nnodes == 1 ? "thread" : "threads"
955 cr->nnodes == 1 ? "process" : "processes"
961 /* Check and update hw_opt for the cut-off scheme */
962 check_and_update_hw_opt_2(&hw_opt, inputrec->cutoff_scheme);
964 /* Check and update the number of OpenMP threads requested */
965 checkAndUpdateRequestedNumOpenmpThreads(&hw_opt, *hwinfo, cr, ms, physicalNodeComm.size_,
968 gmx_omp_nthreads_init(mdlog, cr,
969 hwinfo->nthreads_hw_avail,
970 physicalNodeComm.size_,
972 hw_opt.nthreads_omp_pme,
973 !thisRankHasDuty(cr, DUTY_PP),
974 inputrec->cutoff_scheme == ecutsVERLET);
976 // Enable FP exception detection for the Verlet scheme, but not in
977 // Release mode and not for compilers with known buggy FP
978 // exception support (clang with any optimization) or suspected
979 // buggy FP exception support (gcc 7.* with optimization).
980 #if !defined NDEBUG && \
981 !((defined __clang__ || (defined(__GNUC__) && !defined(__ICC) && __GNUC__ == 7)) \
982 && defined __OPTIMIZE__)
983 const bool bEnableFPE = inputrec->cutoff_scheme == ecutsVERLET;
985 const bool bEnableFPE = false;
987 //FIXME - reconcile with gmx_feenableexcept() call from CommandLineModuleManager::run()
990 gmx_feenableexcept();
993 // Build a data structure that expresses which kinds of non-bonded
994 // task are handled by this rank.
996 // TODO Later, this might become a loop over all registered modules
997 // relevant to the mdp inputs, to find those that have such tasks.
999 // TODO This could move before init_domain_decomposition() as part
1000 // of refactoring that separates the responsibility for duty
1001 // assignment from setup for communication between tasks, and
1002 // setup for tasks handled with a domain (ie including short-ranged
1003 // tasks, bonded tasks, etc.).
1005 // Note that in general useGpuForNonbonded, etc. can have a value
1006 // that is inconsistent with the presence of actual GPUs on any
1007 // rank, and that is not known to be a problem until the
1008 // duty of the ranks on a node become known.
1010 // TODO Later we might need the concept of computeTasksOnThisRank,
1011 // from which we construct gpuTasksOnThisRank.
1013 // Currently the DD code assigns duty to ranks that can
1014 // include PP work that currently can be executed on a single
1015 // GPU, if present and compatible. This has to be coordinated
1016 // across PP ranks on a node, with possible multiple devices
1017 // or sharing devices on a node, either from the user
1018 // selection, or automatically.
1019 auto haveGpus = !gpuIdsToUse.empty();
1020 std::vector<GpuTask> gpuTasksOnThisRank;
1021 if (thisRankHasDuty(cr, DUTY_PP))
1023 if (useGpuForNonbonded)
1025 // Note that any bonded tasks on a GPU always accompany a
1029 gpuTasksOnThisRank.push_back(GpuTask::Nonbonded);
1031 else if (nonbondedTarget == TaskTarget::Gpu)
1033 gmx_fatal(FARGS, "Cannot run short-ranged nonbonded interactions on a GPU because there is none detected.");
1035 else if (bondedTarget == TaskTarget::Gpu)
1037 gmx_fatal(FARGS, "Cannot run bonded interactions on a GPU because there is none detected.");
1041 // TODO cr->duty & DUTY_PME should imply that a PME algorithm is active, but currently does not.
1042 if (EEL_PME(inputrec->coulombtype) && (thisRankHasDuty(cr, DUTY_PME)))
1048 gpuTasksOnThisRank.push_back(GpuTask::Pme);
1050 else if (pmeTarget == TaskTarget::Gpu)
1052 gmx_fatal(FARGS, "Cannot run PME on a GPU because there is none detected.");
1057 GpuTaskAssignment gpuTaskAssignment;
1060 // Produce the task assignment for this rank.
1061 gpuTaskAssignment = runTaskAssignment(gpuIdsToUse, userGpuTaskAssignment, *hwinfo,
1062 mdlog, cr, ms, physicalNodeComm, gpuTasksOnThisRank,
1063 useGpuForBonded, pmeRunMode);
1065 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1067 /* Prevent other ranks from continuing after an issue was found
1068 * and reported as a fatal error.
1070 * TODO This function implements a barrier so that MPI runtimes
1071 * can organize an orderly shutdown if one of the ranks has had to
1072 * issue a fatal error in various code already run. When we have
1073 * MPI-aware error handling and reporting, this should be
1078 MPI_Barrier(cr->mpi_comm_mysim);
1084 MPI_Barrier(ms->mpi_comm_masters);
1086 /* We need another barrier to prevent non-master ranks from contiuing
1087 * when an error occured in a different simulation.
1089 MPI_Barrier(cr->mpi_comm_mysim);
1093 /* Now that we know the setup is consistent, check for efficiency */
1094 check_resource_division_efficiency(hwinfo, !gpuTaskAssignment.empty(), mdrunOptions.ntompOptionIsSet,
1097 gmx_device_info_t *nonbondedDeviceInfo = nullptr;
1099 if (thisRankHasDuty(cr, DUTY_PP))
1101 // This works because only one task of each type is currently permitted.
1102 auto nbGpuTaskMapping = std::find_if(gpuTaskAssignment.begin(), gpuTaskAssignment.end(),
1103 hasTaskType<GpuTask::Nonbonded>);
1104 if (nbGpuTaskMapping != gpuTaskAssignment.end())
1106 int nonbondedDeviceId = nbGpuTaskMapping->deviceId_;
1107 nonbondedDeviceInfo = getDeviceInfo(hwinfo->gpu_info, nonbondedDeviceId);
1108 init_gpu(nonbondedDeviceInfo);
1110 if (DOMAINDECOMP(cr))
1112 /* When we share GPUs over ranks, we need to know this for the DLB */
1113 dd_setup_dlb_resource_sharing(cr, nonbondedDeviceId);
1119 std::unique_ptr<ClfftInitializer> initializedClfftLibrary;
1121 gmx_device_info_t *pmeDeviceInfo = nullptr;
1122 // Later, this program could contain kernels that might be later
1123 // re-used as auto-tuning progresses, or subsequent simulations
1125 PmeGpuProgramStorage pmeGpuProgram;
1126 // This works because only one task of each type is currently permitted.
1127 auto pmeGpuTaskMapping = std::find_if(gpuTaskAssignment.begin(), gpuTaskAssignment.end(), hasTaskType<GpuTask::Pme>);
1128 const bool thisRankHasPmeGpuTask = (pmeGpuTaskMapping != gpuTaskAssignment.end());
1129 if (thisRankHasPmeGpuTask)
1131 pmeDeviceInfo = getDeviceInfo(hwinfo->gpu_info, pmeGpuTaskMapping->deviceId_);
1132 init_gpu(pmeDeviceInfo);
1133 pmeGpuProgram = buildPmeGpuProgram(pmeDeviceInfo);
1134 // TODO It would be nice to move this logic into the factory
1135 // function. See Redmine #2535.
1136 bool isMasterThread = !GMX_THREAD_MPI || MASTER(cr);
1137 if (pmeRunMode == PmeRunMode::GPU && !initializedClfftLibrary && isMasterThread)
1139 initializedClfftLibrary = initializeClfftLibrary();
1143 /* getting number of PP/PME threads
1144 PME: env variable should be read only on one node to make sure it is
1145 identical everywhere;
1147 nthreads_pme = gmx_omp_nthreads_get(emntPME);
1149 int numThreadsOnThisRank;
1150 /* threads on this MPI process or TMPI thread */
1151 if (thisRankHasDuty(cr, DUTY_PP))
1153 numThreadsOnThisRank = gmx_omp_nthreads_get(emntNonbonded);
1157 numThreadsOnThisRank = nthreads_pme;
1160 checkHardwareOversubscription(numThreadsOnThisRank, cr->nodeid,
1161 *hwinfo->hardwareTopology,
1162 physicalNodeComm, mdlog);
1164 if (hw_opt.thread_affinity != threadaffOFF)
1166 /* Before setting affinity, check whether the affinity has changed
1167 * - which indicates that probably the OpenMP library has changed it
1168 * since we first checked).
1170 gmx_check_thread_affinity_set(mdlog, cr,
1171 &hw_opt, hwinfo->nthreads_hw_avail, TRUE);
1173 int numThreadsOnThisNode, intraNodeThreadOffset;
1174 analyzeThreadsOnThisNode(physicalNodeComm, numThreadsOnThisRank, &numThreadsOnThisNode,
1175 &intraNodeThreadOffset);
1177 /* Set the CPU affinity */
1178 gmx_set_thread_affinity(mdlog, cr, &hw_opt, *hwinfo->hardwareTopology,
1179 numThreadsOnThisRank, numThreadsOnThisNode,
1180 intraNodeThreadOffset, nullptr);
1183 if (mdrunOptions.timingOptions.resetStep > -1)
1185 GMX_LOG(mdlog.info).asParagraph().
1186 appendText("The -resetstep functionality is deprecated, and may be removed in a future version.");
1188 wcycle = wallcycle_init(fplog, mdrunOptions.timingOptions.resetStep, cr);
1192 /* Master synchronizes its value of reset_counters with all nodes
1193 * including PME only nodes */
1194 reset_counters = wcycle_get_reset_counters(wcycle);
1195 gmx_bcast_sim(sizeof(reset_counters), &reset_counters, cr);
1196 wcycle_set_reset_counters(wcycle, reset_counters);
1199 // Membrane embedding must be initialized before we call init_forcerec()
1204 fprintf(stderr, "Initializing membed");
1206 /* Note that membed cannot work in parallel because mtop is
1207 * changed here. Fix this if we ever want to make it run with
1208 * multiple ranks. */
1209 membed = init_membed(fplog, filenames.size(), filenames.data(), &mtop, inputrec, globalState.get(), cr,
1211 .checkpointOptions.period);
1214 std::unique_ptr<MDAtoms> mdAtoms;
1215 std::unique_ptr<gmx_vsite_t> vsite;
1218 if (thisRankHasDuty(cr, DUTY_PP))
1220 /* Initiate forcerecord */
1222 fr->forceProviders = mdModules->initForceProviders();
1223 init_forcerec(fplog, mdlog, fr, fcd,
1224 inputrec, &mtop, cr, box,
1225 opt2fn("-table", filenames.size(), filenames.data()),
1226 opt2fn("-tablep", filenames.size(), filenames.data()),
1227 opt2fns("-tableb", filenames.size(), filenames.data()),
1228 *hwinfo, nonbondedDeviceInfo,
1233 /* Initialize the mdAtoms structure.
1234 * mdAtoms is not filled with atom data,
1235 * as this can not be done now with domain decomposition.
1237 mdAtoms = makeMDAtoms(fplog, mtop, *inputrec, thisRankHasPmeGpuTask);
1238 if (globalState && thisRankHasPmeGpuTask)
1240 // The pinning of coordinates in the global state object works, because we only use
1241 // PME on GPU without DD or on a separate PME rank, and because the local state pointer
1242 // points to the global state object without DD.
1243 // FIXME: MD and EM separately set up the local state - this should happen in the same function,
1244 // which should also perform the pinning.
1245 changePinningPolicy(&globalState->x, pme_get_pinning_policy());
1248 /* Initialize the virtual site communication */
1249 vsite = initVsite(mtop, cr);
1251 calc_shifts(box, fr->shift_vec);
1253 /* With periodic molecules the charge groups should be whole at start up
1254 * and the virtual sites should not be far from their proper positions.
1256 if (!inputrec->bContinuation && MASTER(cr) &&
1257 !(inputrec->ePBC != epbcNONE && inputrec->bPeriodicMols))
1259 /* Make molecules whole at start of run */
1260 if (fr->ePBC != epbcNONE)
1262 do_pbc_first_mtop(fplog, inputrec->ePBC, box, &mtop, globalState->x.rvec_array());
1266 /* Correct initial vsite positions are required
1267 * for the initial distribution in the domain decomposition
1268 * and for the initial shell prediction.
1270 constructVsitesGlobal(mtop, globalState->x);
1274 if (EEL_PME(fr->ic->eeltype) || EVDW_PME(fr->ic->vdwtype))
1276 ewaldcoeff_q = fr->ic->ewaldcoeff_q;
1277 ewaldcoeff_lj = fr->ic->ewaldcoeff_lj;
1282 /* This is a PME only node */
1284 GMX_ASSERT(globalState == nullptr, "We don't need the state on a PME only rank and expect it to be unitialized");
1286 ewaldcoeff_q = calc_ewaldcoeff_q(inputrec->rcoulomb, inputrec->ewald_rtol);
1287 ewaldcoeff_lj = calc_ewaldcoeff_lj(inputrec->rvdw, inputrec->ewald_rtol_lj);
1290 gmx_pme_t *sepPmeData = nullptr;
1291 // This reference hides the fact that PME data is owned by runner on PME-only ranks and by forcerec on other ranks
1292 GMX_ASSERT(thisRankHasDuty(cr, DUTY_PP) == (fr != nullptr), "Double-checking that only PME-only ranks have no forcerec");
1293 gmx_pme_t * &pmedata = fr ? fr->pmedata : sepPmeData;
1295 /* Initiate PME if necessary,
1296 * either on all nodes or on dedicated PME nodes only. */
1297 if (EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype))
1299 if (mdAtoms && mdAtoms->mdatoms())
1301 nChargePerturbed = mdAtoms->mdatoms()->nChargePerturbed;
1302 if (EVDW_PME(inputrec->vdwtype))
1304 nTypePerturbed = mdAtoms->mdatoms()->nTypePerturbed;
1307 if (cr->npmenodes > 0)
1309 /* The PME only nodes need to know nChargePerturbed(FEP on Q) and nTypePerturbed(FEP on LJ)*/
1310 gmx_bcast_sim(sizeof(nChargePerturbed), &nChargePerturbed, cr);
1311 gmx_bcast_sim(sizeof(nTypePerturbed), &nTypePerturbed, cr);
1314 if (thisRankHasDuty(cr, DUTY_PME))
1318 pmedata = gmx_pme_init(cr,
1319 getNumPmeDomains(cr->dd),
1321 mtop.natoms, nChargePerturbed != 0, nTypePerturbed != 0,
1322 mdrunOptions.reproducible,
1323 ewaldcoeff_q, ewaldcoeff_lj,
1325 pmeRunMode, nullptr,
1326 pmeDeviceInfo, pmeGpuProgram.get(), mdlog);
1328 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1333 if (EI_DYNAMICS(inputrec->eI))
1335 /* Turn on signal handling on all nodes */
1337 * (A user signal from the PME nodes (if any)
1338 * is communicated to the PP nodes.
1340 signal_handler_install();
1343 if (thisRankHasDuty(cr, DUTY_PP))
1345 /* Assumes uniform use of the number of OpenMP threads */
1346 walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(emntDefault));
1348 if (inputrec->bPull)
1350 /* Initialize pull code */
1351 inputrec->pull_work =
1352 init_pull(fplog, inputrec->pull, inputrec,
1353 &mtop, cr, &atomSets, inputrec->fepvals->init_lambda);
1354 if (inputrec->pull->bXOutAverage || inputrec->pull->bFOutAverage)
1356 initPullHistory(inputrec->pull_work, &observablesHistory);
1358 if (EI_DYNAMICS(inputrec->eI) && MASTER(cr))
1360 init_pull_output_files(inputrec->pull_work,
1361 filenames.size(), filenames.data(), oenv,
1362 continuationOptions);
1366 std::unique_ptr<EnforcedRotation> enforcedRotation;
1369 /* Initialize enforced rotation code */
1370 enforcedRotation = init_rot(fplog,
1382 if (inputrec->eSwapCoords != eswapNO)
1384 /* Initialize ion swapping code */
1385 init_swapcoords(fplog, inputrec, opt2fn_master("-swap", filenames.size(), filenames.data(), cr),
1386 &mtop, globalState.get(), &observablesHistory,
1387 cr, &atomSets, oenv, mdrunOptions);
1390 /* Let makeConstraints know whether we have essential dynamics constraints.
1391 * TODO: inputrec should tell us whether we use an algorithm, not a file option or the checkpoint
1393 bool doEssentialDynamics = (opt2fn_null("-ei", filenames.size(), filenames.data()) != nullptr
1394 || observablesHistory.edsamHistory);
1395 auto constr = makeConstraints(mtop, *inputrec, doEssentialDynamics,
1396 fplog, *mdAtoms->mdatoms(),
1397 cr, ms, nrnb, wcycle, fr->bMolPBC);
1399 if (DOMAINDECOMP(cr))
1401 GMX_RELEASE_ASSERT(fr, "fr was NULL while cr->duty was DUTY_PP");
1402 /* This call is not included in init_domain_decomposition mainly
1403 * because fr->cginfo_mb is set later.
1405 dd_init_bondeds(fplog, cr->dd, &mtop, vsite.get(), inputrec,
1406 domdecOptions.checkBondedInteractions,
1410 // TODO This is not the right place to manage the lifetime of
1411 // this data structure, but currently it's the easiest way to
1412 // make it work. Later, it should probably be made/updated
1413 // after the workload for the lifetime of a PP domain is
1415 PpForceWorkload ppForceWorkload;
1417 GMX_ASSERT(stopHandlerBuilder_, "Runner must provide StopHandlerBuilder to integrator.");
1418 /* Now do whatever the user wants us to do (how flexible...) */
1419 Integrator integrator {
1420 fplog, cr, ms, mdlog, static_cast<int>(filenames.size()), filenames.data(),
1423 vsite.get(), constr.get(),
1424 enforcedRotation ? enforcedRotation->getLegacyEnfrot() : nullptr,
1426 mdModules->outputProvider(),
1430 &observablesHistory,
1431 mdAtoms.get(), nrnb, wcycle, fr,
1435 walltime_accounting,
1436 std::move(stopHandlerBuilder_)
1438 integrator.run(inputrec->eI, doRerun);
1440 if (inputrec->bPull)
1442 finish_pull(inputrec->pull_work);
1448 GMX_RELEASE_ASSERT(pmedata, "pmedata was NULL while cr->duty was not DUTY_PP");
1450 walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(emntPME));
1451 gmx_pmeonly(pmedata, cr, nrnb, wcycle, walltime_accounting, inputrec, pmeRunMode);
1454 wallcycle_stop(wcycle, ewcRUN);
1456 /* Finish up, write some stuff
1457 * if rerunMD, don't write last frame again
1459 finish_run(fplog, mdlog, cr,
1460 inputrec, nrnb, wcycle, walltime_accounting,
1461 fr ? fr->nbv : nullptr,
1463 EI_DYNAMICS(inputrec->eI) && !isMultiSim(ms));
1465 // clean up cycle counter
1466 wallcycle_destroy(wcycle);
1471 gmx_pme_destroy(pmedata);
1475 // FIXME: this is only here to manually unpin mdAtoms->chargeA_ and state->x,
1476 // before we destroy the GPU context(s) in free_gpu_resources().
1477 // Pinned buffers are associated with contexts in CUDA.
1478 // As soon as we destroy GPU contexts after mdrunner() exits, these lines should go.
1479 mdAtoms.reset(nullptr);
1480 globalState.reset(nullptr);
1481 mdModules.reset(nullptr); // destruct force providers here as they might also use the GPU
1483 /* Free GPU memory and set a physical node tMPI barrier (which should eventually go away) */
1484 free_gpu_resources(fr, physicalNodeComm);
1485 free_gpu(nonbondedDeviceInfo);
1486 free_gpu(pmeDeviceInfo);
1487 done_forcerec(fr, mtop.molblock.size(), mtop.groups.grps[egcENER].nr);
1492 free_membed(membed);
1495 gmx_hardware_info_free();
1497 /* Does what it says */
1498 print_date_and_time(fplog, cr->nodeid, "Finished mdrun", gmx_gettime());
1499 walltime_accounting_destroy(walltime_accounting);
1502 // Ensure log file content is written
1505 gmx_fio_flush(logFileHandle);
1508 /* Reset FPEs (important for unit tests) by disabling them. Assumes no
1509 * exceptions were enabled before function was called. */
1512 gmx_fedisableexcept();
1515 rc = static_cast<int>(gmx_get_stop_condition());
1518 /* we need to join all threads. The sub-threads join when they
1519 exit this function, but the master thread needs to be told to
1521 if (PAR(cr) && MASTER(cr))
1525 //TODO free commrec in MPI simulations
1531 Mdrunner::~Mdrunner()
1533 // Clean up of the Manager.
1534 // This will end up getting called on every thread-MPI rank, which is unnecessary,
1535 // but okay as long as threads synchronize some time before adding or accessing
1536 // a new set of restraints.
1537 if (restraintManager_)
1539 restraintManager_->clear();
1540 GMX_ASSERT(restraintManager_->countRestraints() == 0,
1541 "restraints added during runner life time should be cleared at runner destruction.");
1545 void Mdrunner::addPotential(std::shared_ptr<gmx::IRestraintPotential> puller,
1548 GMX_ASSERT(restraintManager_, "Mdrunner must have a restraint manager.");
1549 // Not sure if this should be logged through the md logger or something else,
1550 // but it is helpful to have some sort of INFO level message sent somewhere.
1551 // std::cout << "Registering restraint named " << name << std::endl;
1553 // When multiple restraints are used, it may be wasteful to register them separately.
1554 // Maybe instead register an entire Restraint Manager as a force provider.
1555 restraintManager_->addToSpec(std::move(puller),
1559 Mdrunner::Mdrunner(Mdrunner &&) noexcept = default;
1561 //NOLINTNEXTLINE(performance-noexcept-move-constructor) working around GCC bug 58265
1562 Mdrunner &Mdrunner::operator=(Mdrunner && /*handle*/) noexcept(BUGFREE_NOEXCEPT_STRING) = default;
1564 class Mdrunner::BuilderImplementation
1567 BuilderImplementation() = delete;
1568 explicit BuilderImplementation(SimulationContext* context);
1569 ~BuilderImplementation();
1571 BuilderImplementation &setExtraMdrunOptions(const MdrunOptions &options,
1572 real forceWarningThreshold);
1574 void addDomdec(const DomdecOptions &options);
1576 void addVerletList(int nstlist);
1578 void addReplicaExchange(const ReplicaExchangeParameters ¶ms);
1580 void addMultiSim(gmx_multisim_t* multisim);
1582 void addNonBonded(const char* nbpu_opt);
1584 void addPME(const char* pme_opt_, const char* pme_fft_opt_);
1586 void addBondedTaskAssignment(const char* bonded_opt);
1588 void addHardwareOptions(const gmx_hw_opt_t &hardwareOptions);
1590 void addFilenames(ArrayRef <const t_filenm> filenames);
1592 void addOutputEnvironment(gmx_output_env_t* outputEnvironment);
1594 void addLogFile(t_fileio *logFileHandle);
1596 void addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder);
1601 // Default parameters copied from runner.h
1602 // \todo Clarify source(s) of default parameters.
1604 const char* nbpu_opt_ = nullptr;
1605 const char* pme_opt_ = nullptr;
1606 const char* pme_fft_opt_ = nullptr;
1607 const char *bonded_opt_ = nullptr;
1609 MdrunOptions mdrunOptions_;
1611 DomdecOptions domdecOptions_;
1613 ReplicaExchangeParameters replicaExchangeParameters_;
1615 //! Command-line override for the duration of a neighbor list with the Verlet scheme.
1618 //! Non-owning multisim communicator handle.
1619 std::unique_ptr<gmx_multisim_t*> multisim_ = nullptr;
1621 //! Print a warning if any force is larger than this (in kJ/mol nm).
1622 real forceWarningThreshold_ = -1;
1624 /*! \brief Non-owning pointer to SimulationContext (owned and managed by client)
1627 * \todo Establish robust protocol to make sure resources remain valid.
1628 * SimulationContext will likely be separated into multiple layers for
1629 * different levels of access and different phases of execution. Ref
1630 * https://redmine.gromacs.org/issues/2375
1631 * https://redmine.gromacs.org/issues/2587
1633 SimulationContext* context_ = nullptr;
1635 //! \brief Parallelism information.
1636 gmx_hw_opt_t hardwareOptions_;
1638 //! filename options for simulation.
1639 ArrayRef<const t_filenm> filenames_;
1641 /*! \brief Handle to output environment.
1643 * \todo gmx_output_env_t needs lifetime management.
1645 gmx_output_env_t* outputEnvironment_ = nullptr;
1647 /*! \brief Non-owning handle to MD log file.
1649 * \todo Context should own output facilities for client.
1650 * \todo Improve log file handle management.
1652 * Code managing the FILE* relies on the ability to set it to
1653 * nullptr to check whether the filehandle is valid.
1655 t_fileio* logFileHandle_ = nullptr;
1658 * \brief Builder for simulation stop signal handler.
1660 std::unique_ptr<StopHandlerBuilder> stopHandlerBuilder_ = nullptr;
1663 Mdrunner::BuilderImplementation::BuilderImplementation(SimulationContext* context) :
1666 GMX_ASSERT(context_, "Bug found. It should not be possible to construct builder without a valid context.");
1669 Mdrunner::BuilderImplementation::~BuilderImplementation() = default;
1671 Mdrunner::BuilderImplementation &
1672 Mdrunner::BuilderImplementation::setExtraMdrunOptions(const MdrunOptions &options,
1673 real forceWarningThreshold)
1675 mdrunOptions_ = options;
1676 forceWarningThreshold_ = forceWarningThreshold;
1680 void Mdrunner::BuilderImplementation::addDomdec(const DomdecOptions &options)
1682 domdecOptions_ = options;
1685 void Mdrunner::BuilderImplementation::addVerletList(int nstlist)
1690 void Mdrunner::BuilderImplementation::addReplicaExchange(const ReplicaExchangeParameters ¶ms)
1692 replicaExchangeParameters_ = params;
1695 void Mdrunner::BuilderImplementation::addMultiSim(gmx_multisim_t* multisim)
1697 multisim_ = compat::make_unique<gmx_multisim_t*>(multisim);
1700 Mdrunner Mdrunner::BuilderImplementation::build()
1702 auto newRunner = Mdrunner();
1704 GMX_ASSERT(context_, "Bug found. It should not be possible to call build() without a valid context.");
1706 newRunner.mdrunOptions = mdrunOptions_;
1707 newRunner.domdecOptions = domdecOptions_;
1709 // \todo determine an invariant to check or confirm that all gmx_hw_opt_t objects are valid
1710 newRunner.hw_opt = hardwareOptions_;
1712 // No invariant to check. This parameter exists to optionally override other behavior.
1713 newRunner.nstlist_cmdline = nstlist_;
1715 newRunner.replExParams = replicaExchangeParameters_;
1717 newRunner.filenames = filenames_;
1719 GMX_ASSERT(context_->communicationRecord_, "SimulationContext communications not initialized.");
1720 newRunner.cr = context_->communicationRecord_;
1724 // nullptr is a valid value for the multisim handle, so we don't check the pointed-to pointer.
1725 newRunner.ms = *multisim_;
1729 GMX_THROW(gmx::APIError("MdrunnerBuilder::addMultiSim() is required before build()"));
1732 // \todo Clarify ownership and lifetime management for gmx_output_env_t
1733 // \todo Update sanity checking when output environment has clearly specified invariants.
1734 // Initialization and default values for oenv are not well specified in the current version.
1735 if (outputEnvironment_)
1737 newRunner.oenv = outputEnvironment_;
1741 GMX_THROW(gmx::APIError("MdrunnerBuilder::addOutputEnvironment() is required before build()"));
1744 newRunner.logFileHandle = logFileHandle_;
1748 newRunner.nbpu_opt = nbpu_opt_;
1752 GMX_THROW(gmx::APIError("MdrunnerBuilder::addNonBonded() is required before build()"));
1755 if (pme_opt_ && pme_fft_opt_)
1757 newRunner.pme_opt = pme_opt_;
1758 newRunner.pme_fft_opt = pme_fft_opt_;
1762 GMX_THROW(gmx::APIError("MdrunnerBuilder::addElectrostatics() is required before build()"));
1767 newRunner.bonded_opt = bonded_opt_;
1771 GMX_THROW(gmx::APIError("MdrunnerBuilder::addBondedTaskAssignment() is required before build()"));
1774 newRunner.restraintManager_ = compat::make_unique<gmx::RestraintManager>();
1776 if (stopHandlerBuilder_)
1778 newRunner.stopHandlerBuilder_ = std::move(stopHandlerBuilder_);
1782 newRunner.stopHandlerBuilder_ = compat::make_unique<StopHandlerBuilder>();
1788 void Mdrunner::BuilderImplementation::addNonBonded(const char* nbpu_opt)
1790 nbpu_opt_ = nbpu_opt;
1793 void Mdrunner::BuilderImplementation::addPME(const char* pme_opt,
1794 const char* pme_fft_opt)
1797 pme_fft_opt_ = pme_fft_opt;
1800 void Mdrunner::BuilderImplementation::addBondedTaskAssignment(const char* bonded_opt)
1802 bonded_opt_ = bonded_opt;
1805 void Mdrunner::BuilderImplementation::addHardwareOptions(const gmx_hw_opt_t &hardwareOptions)
1807 hardwareOptions_ = hardwareOptions;
1810 void Mdrunner::BuilderImplementation::addFilenames(ArrayRef<const t_filenm> filenames)
1812 filenames_ = filenames;
1815 void Mdrunner::BuilderImplementation::addOutputEnvironment(gmx_output_env_t* outputEnvironment)
1817 outputEnvironment_ = outputEnvironment;
1820 void Mdrunner::BuilderImplementation::addLogFile(t_fileio *logFileHandle)
1822 logFileHandle_ = logFileHandle;
1825 void Mdrunner::BuilderImplementation::addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder)
1827 stopHandlerBuilder_ = std::move(builder);
1830 MdrunnerBuilder::MdrunnerBuilder(compat::not_null<SimulationContext*> context) :
1831 impl_ {gmx::compat::make_unique<Mdrunner::BuilderImplementation>(context)}
1835 MdrunnerBuilder::~MdrunnerBuilder() = default;
1837 MdrunnerBuilder &MdrunnerBuilder::addSimulationMethod(const MdrunOptions &options,
1838 real forceWarningThreshold)
1840 impl_->setExtraMdrunOptions(options, forceWarningThreshold);
1844 MdrunnerBuilder &MdrunnerBuilder::addDomainDecomposition(const DomdecOptions &options)
1846 impl_->addDomdec(options);
1850 MdrunnerBuilder &MdrunnerBuilder::addNeighborList(int nstlist)
1852 impl_->addVerletList(nstlist);
1856 MdrunnerBuilder &MdrunnerBuilder::addReplicaExchange(const ReplicaExchangeParameters ¶ms)
1858 impl_->addReplicaExchange(params);
1862 MdrunnerBuilder &MdrunnerBuilder::addMultiSim(gmx_multisim_t* multisim)
1864 impl_->addMultiSim(multisim);
1868 MdrunnerBuilder &MdrunnerBuilder::addNonBonded(const char* nbpu_opt)
1870 impl_->addNonBonded(nbpu_opt);
1874 MdrunnerBuilder &MdrunnerBuilder::addElectrostatics(const char* pme_opt,
1875 const char* pme_fft_opt)
1877 // The builder method may become more general in the future, but in this version,
1878 // parameters for PME electrostatics are both required and the only parameters
1880 if (pme_opt && pme_fft_opt)
1882 impl_->addPME(pme_opt, pme_fft_opt);
1886 GMX_THROW(gmx::InvalidInputError("addElectrostatics() arguments must be non-null pointers."));
1891 MdrunnerBuilder &MdrunnerBuilder::addBondedTaskAssignment(const char* bonded_opt)
1893 impl_->addBondedTaskAssignment(bonded_opt);
1897 Mdrunner MdrunnerBuilder::build()
1899 return impl_->build();
1902 MdrunnerBuilder &MdrunnerBuilder::addHardwareOptions(const gmx_hw_opt_t &hardwareOptions)
1904 impl_->addHardwareOptions(hardwareOptions);
1908 MdrunnerBuilder &MdrunnerBuilder::addFilenames(ArrayRef<const t_filenm> filenames)
1910 impl_->addFilenames(filenames);
1914 MdrunnerBuilder &MdrunnerBuilder::addOutputEnvironment(gmx_output_env_t* outputEnvironment)
1916 impl_->addOutputEnvironment(outputEnvironment);
1920 MdrunnerBuilder &MdrunnerBuilder::addLogFile(t_fileio *logFileHandle)
1922 impl_->addLogFile(logFileHandle);
1926 MdrunnerBuilder &MdrunnerBuilder::addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder)
1928 impl_->addStopHandlerBuilder(std::move(builder));
1932 MdrunnerBuilder::MdrunnerBuilder(MdrunnerBuilder &&) noexcept = default;
1934 MdrunnerBuilder &MdrunnerBuilder::operator=(MdrunnerBuilder &&) noexcept = default;