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, by the GROMACS development team, led by
7 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
8 * and including many others, as listed in the AUTHORS file in the
9 * top-level source directory and at http://www.gromacs.org.
11 * GROMACS is free software; you can redistribute it and/or
12 * modify it under the terms of the GNU Lesser General Public License
13 * as published by the Free Software Foundation; either version 2.1
14 * of the License, or (at your option) any later version.
16 * GROMACS is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
19 * Lesser General Public License for more details.
21 * You should have received a copy of the GNU Lesser General Public
22 * License along with GROMACS; if not, see
23 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
24 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
26 * If you want to redistribute modifications to GROMACS, please
27 * consider that scientific software is very special. Version
28 * control is crucial - bugs must be traceable. We will be happy to
29 * consider code for inclusion in the official distribution, but
30 * derived work must not be called official GROMACS. Details are found
31 * in the README & COPYING files - if they are missing, get the
32 * official version at http://www.gromacs.org.
34 * To help us fund GROMACS development, we humbly ask that you cite
35 * the research papers on the package. Check out http://www.gromacs.org.
39 * \brief Implements the MD runner routine calling all integrators.
41 * \author David van der Spoel <david.vanderspoel@icm.uu.se>
42 * \ingroup module_mdrun
59 #include "gromacs/commandline/filenm.h"
60 #include "gromacs/domdec/builder.h"
61 #include "gromacs/domdec/domdec.h"
62 #include "gromacs/domdec/domdec_struct.h"
63 #include "gromacs/domdec/gpuhaloexchange.h"
64 #include "gromacs/domdec/localatomsetmanager.h"
65 #include "gromacs/domdec/partition.h"
66 #include "gromacs/ewald/ewald_utils.h"
67 #include "gromacs/ewald/pme_gpu_program.h"
68 #include "gromacs/ewald/pme_only.h"
69 #include "gromacs/ewald/pme_pp_comm_gpu.h"
70 #include "gromacs/fileio/checkpoint.h"
71 #include "gromacs/fileio/gmxfio.h"
72 #include "gromacs/fileio/oenv.h"
73 #include "gromacs/fileio/tpxio.h"
74 #include "gromacs/gmxlib/network.h"
75 #include "gromacs/gmxlib/nrnb.h"
76 #include "gromacs/gpu_utils/gpu_utils.h"
77 #include "gromacs/hardware/cpuinfo.h"
78 #include "gromacs/hardware/detecthardware.h"
79 #include "gromacs/hardware/printhardware.h"
80 #include "gromacs/imd/imd.h"
81 #include "gromacs/listed_forces/disre.h"
82 #include "gromacs/listed_forces/gpubonded.h"
83 #include "gromacs/listed_forces/orires.h"
84 #include "gromacs/math/functions.h"
85 #include "gromacs/math/utilities.h"
86 #include "gromacs/math/vec.h"
87 #include "gromacs/mdlib/boxdeformation.h"
88 #include "gromacs/mdlib/broadcaststructs.h"
89 #include "gromacs/mdlib/calc_verletbuf.h"
90 #include "gromacs/mdlib/dispersioncorrection.h"
91 #include "gromacs/mdlib/enerdata_utils.h"
92 #include "gromacs/mdlib/force.h"
93 #include "gromacs/mdlib/forcerec.h"
94 #include "gromacs/mdlib/gmx_omp_nthreads.h"
95 #include "gromacs/mdlib/makeconstraints.h"
96 #include "gromacs/mdlib/md_support.h"
97 #include "gromacs/mdlib/mdatoms.h"
98 #include "gromacs/mdlib/membed.h"
99 #include "gromacs/mdlib/qmmm.h"
100 #include "gromacs/mdlib/sighandler.h"
101 #include "gromacs/mdlib/stophandler.h"
102 #include "gromacs/mdlib/updategroups.h"
103 #include "gromacs/mdrun/mdmodules.h"
104 #include "gromacs/mdrun/simulationcontext.h"
105 #include "gromacs/mdrunutility/handlerestart.h"
106 #include "gromacs/mdrunutility/logging.h"
107 #include "gromacs/mdrunutility/multisim.h"
108 #include "gromacs/mdrunutility/printtime.h"
109 #include "gromacs/mdrunutility/threadaffinity.h"
110 #include "gromacs/mdtypes/commrec.h"
111 #include "gromacs/mdtypes/enerdata.h"
112 #include "gromacs/mdtypes/fcdata.h"
113 #include "gromacs/mdtypes/group.h"
114 #include "gromacs/mdtypes/inputrec.h"
115 #include "gromacs/mdtypes/md_enums.h"
116 #include "gromacs/mdtypes/mdrunoptions.h"
117 #include "gromacs/mdtypes/observableshistory.h"
118 #include "gromacs/mdtypes/simulation_workload.h"
119 #include "gromacs/mdtypes/state.h"
120 #include "gromacs/mdtypes/state_propagator_data_gpu.h"
121 #include "gromacs/nbnxm/gpu_data_mgmt.h"
122 #include "gromacs/nbnxm/nbnxm.h"
123 #include "gromacs/nbnxm/pairlist_tuning.h"
124 #include "gromacs/pbcutil/pbc.h"
125 #include "gromacs/pulling/output.h"
126 #include "gromacs/pulling/pull.h"
127 #include "gromacs/pulling/pull_rotation.h"
128 #include "gromacs/restraint/manager.h"
129 #include "gromacs/restraint/restraintmdmodule.h"
130 #include "gromacs/restraint/restraintpotential.h"
131 #include "gromacs/swap/swapcoords.h"
132 #include "gromacs/taskassignment/decidegpuusage.h"
133 #include "gromacs/taskassignment/decidesimulationworkload.h"
134 #include "gromacs/taskassignment/resourcedivision.h"
135 #include "gromacs/taskassignment/taskassignment.h"
136 #include "gromacs/taskassignment/usergpuids.h"
137 #include "gromacs/timing/gpu_timing.h"
138 #include "gromacs/timing/wallcycle.h"
139 #include "gromacs/timing/wallcyclereporting.h"
140 #include "gromacs/topology/mtop_util.h"
141 #include "gromacs/trajectory/trajectoryframe.h"
142 #include "gromacs/utility/basenetwork.h"
143 #include "gromacs/utility/cstringutil.h"
144 #include "gromacs/utility/exceptions.h"
145 #include "gromacs/utility/fatalerror.h"
146 #include "gromacs/utility/filestream.h"
147 #include "gromacs/utility/gmxassert.h"
148 #include "gromacs/utility/gmxmpi.h"
149 #include "gromacs/utility/keyvaluetree.h"
150 #include "gromacs/utility/logger.h"
151 #include "gromacs/utility/loggerbuilder.h"
152 #include "gromacs/utility/mdmodulenotification.h"
153 #include "gromacs/utility/physicalnodecommunicator.h"
154 #include "gromacs/utility/pleasecite.h"
155 #include "gromacs/utility/programcontext.h"
156 #include "gromacs/utility/smalloc.h"
157 #include "gromacs/utility/stringutil.h"
159 #include "isimulator.h"
160 #include "replicaexchange.h"
161 #include "simulatorbuilder.h"
164 # include "corewrap.h"
170 /*! \brief Structure that holds boolean flags corresponding to the development
171 * features present enabled through environment variables.
174 struct DevelopmentFeatureFlags
176 //! True if the Buffer ops development feature is enabled
177 // TODO: when the trigger of the buffer ops offload is fully automated this should go away
178 bool enableGpuBufferOps
= false;
179 //! If true, forces 'mdrun -update auto' default to 'gpu'
180 bool forceGpuUpdateDefault
= false;
181 //! True if the GPU halo exchange development feature is enabled
182 bool enableGpuHaloExchange
= false;
183 //! True if the PME PP direct communication GPU development feature is enabled
184 bool enableGpuPmePPComm
= false;
187 /*! \brief Manage any development feature flag variables encountered
189 * The use of dev features indicated by environment variables is
190 * logged in order to ensure that runs with such features enabled can
191 * be identified from their log and standard output. Any cross
192 * dependencies are also checked, and if unsatisfied, a fatal error
195 * Note that some development features overrides are applied already here:
196 * the GPU communication flags are set to false in non-tMPI and non-CUDA builds.
198 * \param[in] mdlog Logger object.
199 * \param[in] useGpuForNonbonded True if the nonbonded task is offloaded in this run.
200 * \param[in] pmeRunMode The PME run mode for this run
201 * \returns The object populated with development feature flags.
203 static DevelopmentFeatureFlags
manageDevelopmentFeatures(const gmx::MDLogger
& mdlog
,
204 const bool useGpuForNonbonded
,
205 const PmeRunMode pmeRunMode
)
207 DevelopmentFeatureFlags devFlags
;
209 // Some builds of GCC 5 give false positive warnings that these
210 // getenv results are ignored when clearly they are used.
211 #pragma GCC diagnostic push
212 #pragma GCC diagnostic ignored "-Wunused-result"
213 devFlags
.enableGpuBufferOps
= (getenv("GMX_USE_GPU_BUFFER_OPS") != nullptr)
214 && (GMX_GPU
== GMX_GPU_CUDA
) && useGpuForNonbonded
;
215 devFlags
.forceGpuUpdateDefault
= (getenv("GMX_FORCE_UPDATE_DEFAULT_GPU") != nullptr);
216 devFlags
.enableGpuHaloExchange
=
217 (getenv("GMX_GPU_DD_COMMS") != nullptr && GMX_THREAD_MPI
&& (GMX_GPU
== GMX_GPU_CUDA
));
218 devFlags
.enableGpuPmePPComm
=
219 (getenv("GMX_GPU_PME_PP_COMMS") != nullptr && GMX_THREAD_MPI
&& (GMX_GPU
== GMX_GPU_CUDA
));
220 #pragma GCC diagnostic pop
222 if (devFlags
.enableGpuBufferOps
)
224 GMX_LOG(mdlog
.warning
)
226 .appendTextFormatted(
227 "This run uses the 'GPU buffer ops' feature, enabled by the "
228 "GMX_USE_GPU_BUFFER_OPS environment variable.");
231 if (devFlags
.forceGpuUpdateDefault
)
233 GMX_LOG(mdlog
.warning
)
235 .appendTextFormatted(
236 "This run will default to '-update gpu' as requested by the "
237 "GMX_FORCE_UPDATE_DEFAULT_GPU environment variable. GPU update with domain "
238 "decomposition lacks substantial testing and should be used with caution.");
241 if (devFlags
.enableGpuHaloExchange
)
243 if (useGpuForNonbonded
)
245 if (!devFlags
.enableGpuBufferOps
)
247 GMX_LOG(mdlog
.warning
)
249 .appendTextFormatted(
250 "Enabling GPU buffer operations required by GMX_GPU_DD_COMMS "
251 "(equivalent with GMX_USE_GPU_BUFFER_OPS=1).");
252 devFlags
.enableGpuBufferOps
= true;
254 GMX_LOG(mdlog
.warning
)
256 .appendTextFormatted(
257 "This run uses the 'GPU halo exchange' feature, enabled by the "
258 "GMX_GPU_DD_COMMS environment variable.");
262 GMX_LOG(mdlog
.warning
)
264 .appendTextFormatted(
265 "GMX_GPU_DD_COMMS environment variable detected, but the 'GPU "
266 "halo exchange' feature will not be enabled as nonbonded interactions "
267 "are not offloaded.");
268 devFlags
.enableGpuHaloExchange
= false;
272 if (devFlags
.enableGpuPmePPComm
)
274 if (pmeRunMode
== PmeRunMode::GPU
)
276 GMX_LOG(mdlog
.warning
)
278 .appendTextFormatted(
279 "This run uses the 'GPU PME-PP communications' feature, enabled "
280 "by the GMX_GPU_PME_PP_COMMS environment variable.");
284 std::string clarification
;
285 if (pmeRunMode
== PmeRunMode::Mixed
)
288 "PME FFT and gather are not offloaded to the GPU (PME is running in mixed "
293 clarification
= "PME is not offloaded to the GPU.";
295 GMX_LOG(mdlog
.warning
)
298 "GMX_GPU_PME_PP_COMMS environment variable detected, but the "
299 "'GPU PME-PP communications' feature was not enabled as "
301 devFlags
.enableGpuPmePPComm
= false;
308 /*! \brief Barrier for safe simultaneous thread access to mdrunner data
310 * Used to ensure that the master thread does not modify mdrunner during copy
311 * on the spawned threads. */
312 static void threadMpiMdrunnerAccessBarrier()
315 MPI_Barrier(MPI_COMM_WORLD
);
319 Mdrunner
Mdrunner::cloneOnSpawnedThread() const
321 auto newRunner
= Mdrunner(std::make_unique
<MDModules
>());
323 // All runners in the same process share a restraint manager resource because it is
324 // part of the interface to the client code, which is associated only with the
325 // original thread. Handles to the same resources can be obtained by copy.
327 newRunner
.restraintManager_
= std::make_unique
<RestraintManager
>(*restraintManager_
);
330 // Copy members of master runner.
331 // \todo Replace with builder when Simulation context and/or runner phases are better defined.
332 // Ref https://redmine.gromacs.org/issues/2587 and https://redmine.gromacs.org/issues/2375
333 newRunner
.hw_opt
= hw_opt
;
334 newRunner
.filenames
= filenames
;
336 newRunner
.oenv
= oenv
;
337 newRunner
.mdrunOptions
= mdrunOptions
;
338 newRunner
.domdecOptions
= domdecOptions
;
339 newRunner
.nbpu_opt
= nbpu_opt
;
340 newRunner
.pme_opt
= pme_opt
;
341 newRunner
.pme_fft_opt
= pme_fft_opt
;
342 newRunner
.bonded_opt
= bonded_opt
;
343 newRunner
.update_opt
= update_opt
;
344 newRunner
.nstlist_cmdline
= nstlist_cmdline
;
345 newRunner
.replExParams
= replExParams
;
346 newRunner
.pforce
= pforce
;
347 // Give the spawned thread the newly created valid communicator
348 // for the simulation.
349 newRunner
.communicator
= MPI_COMM_WORLD
;
351 newRunner
.startingBehavior
= startingBehavior
;
352 newRunner
.stopHandlerBuilder_
= std::make_unique
<StopHandlerBuilder
>(*stopHandlerBuilder_
);
354 threadMpiMdrunnerAccessBarrier();
359 /*! \brief The callback used for running on spawned threads.
361 * Obtains the pointer to the master mdrunner object from the one
362 * argument permitted to the thread-launch API call, copies it to make
363 * a new runner for this thread, reinitializes necessary data, and
364 * proceeds to the simulation. */
365 static void mdrunner_start_fn(const void* arg
)
369 auto masterMdrunner
= reinterpret_cast<const gmx::Mdrunner
*>(arg
);
370 /* copy the arg list to make sure that it's thread-local. This
371 doesn't copy pointed-to items, of course; fnm, cr and fplog
372 are reset in the call below, all others should be const. */
373 gmx::Mdrunner mdrunner
= masterMdrunner
->cloneOnSpawnedThread();
376 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
380 void Mdrunner::spawnThreads(int numThreadsToLaunch
)
383 /* now spawn new threads that start mdrunner_start_fn(), while
384 the main thread returns. Thread affinity is handled later. */
385 if (tMPI_Init_fn(TRUE
, numThreadsToLaunch
, TMPI_AFFINITY_NONE
, mdrunner_start_fn
,
386 static_cast<const void*>(this))
389 GMX_THROW(gmx::InternalError("Failed to spawn thread-MPI threads"));
392 // Give the master thread the newly created valid communicator for
394 communicator
= MPI_COMM_WORLD
;
395 threadMpiMdrunnerAccessBarrier();
397 GMX_UNUSED_VALUE(numThreadsToLaunch
);
398 GMX_UNUSED_VALUE(mdrunner_start_fn
);
404 /*! \brief Initialize variables for Verlet scheme simulation */
405 static void prepare_verlet_scheme(FILE* fplog
,
409 const gmx_mtop_t
* mtop
,
411 bool makeGpuPairList
,
412 const gmx::CpuInfo
& cpuinfo
)
414 /* For NVE simulations, we will retain the initial list buffer */
415 if (EI_DYNAMICS(ir
->eI
) && ir
->verletbuf_tol
> 0 && !(EI_MD(ir
->eI
) && ir
->etc
== etcNO
))
417 /* Update the Verlet buffer size for the current run setup */
419 /* Here we assume SIMD-enabled kernels are being used. But as currently
420 * calc_verlet_buffer_size gives the same results for 4x8 and 4x4
421 * and 4x2 gives a larger buffer than 4x4, this is ok.
423 ListSetupType listType
=
424 (makeGpuPairList
? ListSetupType::Gpu
: ListSetupType::CpuSimdWhenSupported
);
425 VerletbufListSetup listSetup
= verletbufGetSafeListSetup(listType
);
427 const real rlist_new
=
428 calcVerletBufferSize(*mtop
, det(box
), *ir
, ir
->nstlist
, ir
->nstlist
- 1, -1, listSetup
);
430 if (rlist_new
!= ir
->rlist
)
432 if (fplog
!= nullptr)
435 "\nChanging rlist from %g to %g for non-bonded %dx%d atom kernels\n\n",
436 ir
->rlist
, rlist_new
, listSetup
.cluster_size_i
, listSetup
.cluster_size_j
);
438 ir
->rlist
= rlist_new
;
442 if (nstlist_cmdline
> 0 && (!EI_DYNAMICS(ir
->eI
) || ir
->verletbuf_tol
<= 0))
444 gmx_fatal(FARGS
, "Can not set nstlist without %s",
445 !EI_DYNAMICS(ir
->eI
) ? "dynamics" : "verlet-buffer-tolerance");
448 if (EI_DYNAMICS(ir
->eI
))
450 /* Set or try nstlist values */
451 increaseNstlist(fplog
, cr
, ir
, nstlist_cmdline
, mtop
, box
, makeGpuPairList
, cpuinfo
);
455 /*! \brief Override the nslist value in inputrec
457 * with value passed on the command line (if any)
459 static void override_nsteps_cmdline(const gmx::MDLogger
& mdlog
, int64_t nsteps_cmdline
, t_inputrec
* ir
)
463 /* override with anything else than the default -2 */
464 if (nsteps_cmdline
> -2)
466 char sbuf_steps
[STEPSTRSIZE
];
467 char sbuf_msg
[STRLEN
];
469 ir
->nsteps
= nsteps_cmdline
;
470 if (EI_DYNAMICS(ir
->eI
) && nsteps_cmdline
!= -1)
473 "Overriding nsteps with value passed on the command line: %s steps, %.3g ps",
474 gmx_step_str(nsteps_cmdline
, sbuf_steps
), fabs(nsteps_cmdline
* ir
->delta_t
));
478 sprintf(sbuf_msg
, "Overriding nsteps with value passed on the command line: %s steps",
479 gmx_step_str(nsteps_cmdline
, sbuf_steps
));
482 GMX_LOG(mdlog
.warning
).asParagraph().appendText(sbuf_msg
);
484 else if (nsteps_cmdline
< -2)
486 gmx_fatal(FARGS
, "Invalid nsteps value passed on the command line: %" PRId64
, nsteps_cmdline
);
488 /* Do nothing if nsteps_cmdline == -2 */
494 /*! \brief Return whether GPU acceleration of nonbondeds is supported with the given settings.
496 * If not, and if a warning may be issued, logs a warning about
497 * falling back to CPU code. With thread-MPI, only the first
498 * call to this function should have \c issueWarning true. */
499 static bool gpuAccelerationOfNonbondedIsUseful(const MDLogger
& mdlog
, const t_inputrec
& ir
, bool issueWarning
)
501 bool gpuIsUseful
= true;
504 if (ir
.opts
.ngener
- ir
.nwall
> 1)
506 /* The GPU code does not support more than one energy group.
507 * If the user requested GPUs explicitly, a fatal error is given later.
511 "Multiple energy groups is not implemented for GPUs, falling back to the CPU. "
512 "For better performance, run on the GPU without energy groups and then do "
513 "gmx mdrun -rerun option on the trajectory with an energy group .tpr file.";
519 warning
= "TPI is not implemented for GPUs.";
522 if (!gpuIsUseful
&& issueWarning
)
524 GMX_LOG(mdlog
.warning
).asParagraph().appendText(warning
);
530 //! Initializes the logger for mdrun.
531 static gmx::LoggerOwner
buildLogger(FILE* fplog
, const bool isSimulationMasterRank
)
533 gmx::LoggerBuilder builder
;
534 if (fplog
!= nullptr)
536 builder
.addTargetFile(gmx::MDLogger::LogLevel::Info
, fplog
);
538 if (isSimulationMasterRank
)
540 builder
.addTargetStream(gmx::MDLogger::LogLevel::Warning
, &gmx::TextOutputFile::standardError());
542 return builder
.build();
545 //! Make a TaskTarget from an mdrun argument string.
546 static TaskTarget
findTaskTarget(const char* optionString
)
548 TaskTarget returnValue
= TaskTarget::Auto
;
550 if (strncmp(optionString
, "auto", 3) == 0)
552 returnValue
= TaskTarget::Auto
;
554 else if (strncmp(optionString
, "cpu", 3) == 0)
556 returnValue
= TaskTarget::Cpu
;
558 else if (strncmp(optionString
, "gpu", 3) == 0)
560 returnValue
= TaskTarget::Gpu
;
564 GMX_ASSERT(false, "Option string should have been checked for sanity already");
570 //! Finish run, aggregate data to print performance info.
571 static void finish_run(FILE* fplog
,
572 const gmx::MDLogger
& mdlog
,
574 const t_inputrec
* inputrec
,
576 gmx_wallcycle_t wcycle
,
577 gmx_walltime_accounting_t walltime_accounting
,
578 nonbonded_verlet_t
* nbv
,
579 const gmx_pme_t
* pme
,
583 double nbfs
= 0, mflop
= 0;
584 double elapsed_time
, elapsed_time_over_all_ranks
, elapsed_time_over_all_threads
,
585 elapsed_time_over_all_threads_over_all_ranks
;
586 /* Control whether it is valid to print a report. Only the
587 simulation master may print, but it should not do so if the run
588 terminated e.g. before a scheduled reset step. This is
589 complicated by the fact that PME ranks are unaware of the
590 reason why they were sent a pmerecvqxFINISH. To avoid
591 communication deadlocks, we always do the communication for the
592 report, even if we've decided not to write the report, because
593 how long it takes to finish the run is not important when we've
594 decided not to report on the simulation performance.
596 Further, we only report performance for dynamical integrators,
597 because those are the only ones for which we plan to
598 consider doing any optimizations. */
599 bool printReport
= EI_DYNAMICS(inputrec
->eI
) && SIMMASTER(cr
);
601 if (printReport
&& !walltime_accounting_get_valid_finish(walltime_accounting
))
603 GMX_LOG(mdlog
.warning
)
605 .appendText("Simulation ended prematurely, no performance report will be written.");
610 std::unique_ptr
<t_nrnb
> nrnbTotalStorage
;
613 nrnbTotalStorage
= std::make_unique
<t_nrnb
>();
614 nrnb_tot
= nrnbTotalStorage
.get();
616 MPI_Allreduce(nrnb
->n
, nrnb_tot
->n
, eNRNB
, MPI_DOUBLE
, MPI_SUM
, cr
->mpi_comm_mysim
);
624 elapsed_time
= walltime_accounting_get_time_since_reset(walltime_accounting
);
625 elapsed_time_over_all_threads
=
626 walltime_accounting_get_time_since_reset_over_all_threads(walltime_accounting
);
630 /* reduce elapsed_time over all MPI ranks in the current simulation */
631 MPI_Allreduce(&elapsed_time
, &elapsed_time_over_all_ranks
, 1, MPI_DOUBLE
, MPI_SUM
,
633 elapsed_time_over_all_ranks
/= cr
->nnodes
;
634 /* Reduce elapsed_time_over_all_threads over all MPI ranks in the
635 * current simulation. */
636 MPI_Allreduce(&elapsed_time_over_all_threads
, &elapsed_time_over_all_threads_over_all_ranks
,
637 1, MPI_DOUBLE
, MPI_SUM
, cr
->mpi_comm_mysim
);
642 elapsed_time_over_all_ranks
= elapsed_time
;
643 elapsed_time_over_all_threads_over_all_ranks
= elapsed_time_over_all_threads
;
648 print_flop(fplog
, nrnb_tot
, &nbfs
, &mflop
);
651 if (thisRankHasDuty(cr
, DUTY_PP
) && DOMAINDECOMP(cr
))
653 print_dd_statistics(cr
, inputrec
, fplog
);
656 /* TODO Move the responsibility for any scaling by thread counts
657 * to the code that handled the thread region, so that there's a
658 * mechanism to keep cycle counting working during the transition
659 * to task parallelism. */
660 int nthreads_pp
= gmx_omp_nthreads_get(emntNonbonded
);
661 int nthreads_pme
= gmx_omp_nthreads_get(emntPME
);
662 wallcycle_scale_by_num_threads(wcycle
, thisRankHasDuty(cr
, DUTY_PME
) && !thisRankHasDuty(cr
, DUTY_PP
),
663 nthreads_pp
, nthreads_pme
);
664 auto cycle_sum(wallcycle_sum(cr
, wcycle
));
668 auto nbnxn_gpu_timings
=
669 (nbv
!= nullptr && nbv
->useGpu()) ? Nbnxm::gpu_get_timings(nbv
->gpu_nbv
) : nullptr;
670 gmx_wallclock_gpu_pme_t pme_gpu_timings
= {};
672 if (pme_gpu_task_enabled(pme
))
674 pme_gpu_get_timings(pme
, &pme_gpu_timings
);
676 wallcycle_print(fplog
, mdlog
, cr
->nnodes
, cr
->npmenodes
, nthreads_pp
, nthreads_pme
,
677 elapsed_time_over_all_ranks
, wcycle
, cycle_sum
, nbnxn_gpu_timings
,
680 if (EI_DYNAMICS(inputrec
->eI
))
682 delta_t
= inputrec
->delta_t
;
687 print_perf(fplog
, elapsed_time_over_all_threads_over_all_ranks
, elapsed_time_over_all_ranks
,
688 walltime_accounting_get_nsteps_done_since_reset(walltime_accounting
),
689 delta_t
, nbfs
, mflop
);
693 print_perf(stderr
, elapsed_time_over_all_threads_over_all_ranks
, elapsed_time_over_all_ranks
,
694 walltime_accounting_get_nsteps_done_since_reset(walltime_accounting
),
695 delta_t
, nbfs
, mflop
);
700 int Mdrunner::mdrunner()
703 t_forcerec
* fr
= nullptr;
704 t_fcdata
* fcd
= nullptr;
705 real ewaldcoeff_q
= 0;
706 real ewaldcoeff_lj
= 0;
707 int nChargePerturbed
= -1, nTypePerturbed
= 0;
708 gmx_wallcycle_t wcycle
;
709 gmx_walltime_accounting_t walltime_accounting
= nullptr;
710 gmx_membed_t
* membed
= nullptr;
711 gmx_hw_info_t
* hwinfo
= nullptr;
713 /* CAUTION: threads may be started later on in this function, so
714 cr doesn't reflect the final parallel state right now */
717 /* TODO: inputrec should tell us whether we use an algorithm, not a file option */
718 const bool doEssentialDynamics
= opt2bSet("-ei", filenames
.size(), filenames
.data());
719 const bool doMembed
= opt2bSet("-membed", filenames
.size(), filenames
.data());
720 const bool doRerun
= mdrunOptions
.rerun
;
722 // Handle task-assignment related user options.
723 EmulateGpuNonbonded emulateGpuNonbonded
=
724 (getenv("GMX_EMULATE_GPU") != nullptr ? EmulateGpuNonbonded::Yes
: EmulateGpuNonbonded::No
);
726 std::vector
<int> userGpuTaskAssignment
;
729 userGpuTaskAssignment
= parseUserTaskAssignmentString(hw_opt
.userGpuTaskAssignment
);
731 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
732 auto nonbondedTarget
= findTaskTarget(nbpu_opt
);
733 auto pmeTarget
= findTaskTarget(pme_opt
);
734 auto pmeFftTarget
= findTaskTarget(pme_fft_opt
);
735 auto bondedTarget
= findTaskTarget(bonded_opt
);
736 auto updateTarget
= findTaskTarget(update_opt
);
737 PmeRunMode pmeRunMode
= PmeRunMode::None
;
739 FILE* fplog
= nullptr;
740 // If we are appending, we don't write log output because we need
741 // to check that the old log file matches what the checkpoint file
742 // expects. Otherwise, we should start to write log output now if
743 // there is a file ready for it.
744 if (logFileHandle
!= nullptr && startingBehavior
!= StartingBehavior::RestartWithAppending
)
746 fplog
= gmx_fio_getfp(logFileHandle
);
748 const bool isSimulationMasterRank
= findIsSimulationMasterRank(ms
, communicator
);
749 gmx::LoggerOwner
logOwner(buildLogger(fplog
, isSimulationMasterRank
));
750 gmx::MDLogger
mdlog(logOwner
.logger());
752 // TODO The thread-MPI master rank makes a working
753 // PhysicalNodeCommunicator here, but it gets rebuilt by all ranks
754 // after the threads have been launched. This works because no use
755 // is made of that communicator until after the execution paths
756 // have rejoined. But it is likely that we can improve the way
757 // this is expressed, e.g. by expressly running detection only the
758 // master rank for thread-MPI, rather than relying on the mutex
759 // and reference count.
760 PhysicalNodeCommunicator
physicalNodeComm(communicator
, gmx_physicalnode_id_hash());
761 hwinfo
= gmx_detect_hardware(mdlog
, physicalNodeComm
);
763 gmx_print_detected_hardware(fplog
, isSimulationMasterRank
&& isMasterSim(ms
), mdlog
, hwinfo
);
765 std::vector
<int> gpuIdsToUse
= makeGpuIdsToUse(hwinfo
->gpu_info
, hw_opt
.gpuIdsAvailable
);
767 // Print citation requests after all software/hardware printing
768 pleaseCiteGromacs(fplog
);
770 // TODO Replace this by unique_ptr once t_inputrec is C++
771 t_inputrec inputrecInstance
;
772 t_inputrec
* inputrec
= nullptr;
773 std::unique_ptr
<t_state
> globalState
;
775 auto partialDeserializedTpr
= std::make_unique
<PartialDeserializedTprFile
>();
777 if (isSimulationMasterRank
)
779 /* Only the master rank has the global state */
780 globalState
= std::make_unique
<t_state
>();
782 /* Read (nearly) all data required for the simulation
783 * and keep the partly serialized tpr contents to send to other ranks later
785 *partialDeserializedTpr
= read_tpx_state(ftp2fn(efTPR
, filenames
.size(), filenames
.data()),
786 &inputrecInstance
, globalState
.get(), &mtop
);
787 inputrec
= &inputrecInstance
;
790 /* Check and update the hardware options for internal consistency */
791 checkAndUpdateHardwareOptions(mdlog
, &hw_opt
, isSimulationMasterRank
, domdecOptions
.numPmeRanks
,
794 if (GMX_THREAD_MPI
&& isSimulationMasterRank
)
796 bool useGpuForNonbonded
= false;
797 bool useGpuForPme
= false;
800 GMX_RELEASE_ASSERT(inputrec
!= nullptr, "Keep the compiler happy");
802 // If the user specified the number of ranks, then we must
803 // respect that, but in default mode, we need to allow for
804 // the number of GPUs to choose the number of ranks.
805 auto canUseGpuForNonbonded
= buildSupportsNonbondedOnGpu(nullptr);
806 useGpuForNonbonded
= decideWhetherToUseGpusForNonbondedWithThreadMpi(
807 nonbondedTarget
, gpuIdsToUse
, userGpuTaskAssignment
, emulateGpuNonbonded
,
808 canUseGpuForNonbonded
,
809 gpuAccelerationOfNonbondedIsUseful(mdlog
, *inputrec
, GMX_THREAD_MPI
),
810 hw_opt
.nthreads_tmpi
);
811 useGpuForPme
= decideWhetherToUseGpusForPmeWithThreadMpi(
812 useGpuForNonbonded
, pmeTarget
, gpuIdsToUse
, userGpuTaskAssignment
, *hwinfo
,
813 *inputrec
, mtop
, hw_opt
.nthreads_tmpi
, domdecOptions
.numPmeRanks
);
815 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
817 /* Determine how many thread-MPI ranks to start.
819 * TODO Over-writing the user-supplied value here does
820 * prevent any possible subsequent checks from working
822 hw_opt
.nthreads_tmpi
= get_nthreads_mpi(hwinfo
, &hw_opt
, gpuIdsToUse
, useGpuForNonbonded
,
823 useGpuForPme
, inputrec
, &mtop
, mdlog
, doMembed
);
825 // Now start the threads for thread MPI.
826 spawnThreads(hw_opt
.nthreads_tmpi
);
827 // The spawned threads enter mdrunner() and execution of
828 // master and spawned threads joins at the end of this block.
829 physicalNodeComm
= PhysicalNodeCommunicator(communicator
, gmx_physicalnode_id_hash());
832 GMX_RELEASE_ASSERT(communicator
== MPI_COMM_WORLD
, "Must have valid world communicator");
833 CommrecHandle crHandle
= init_commrec(communicator
, ms
);
834 t_commrec
* cr
= crHandle
.get();
835 GMX_RELEASE_ASSERT(cr
!= nullptr, "Must have valid commrec");
839 /* now broadcast everything to the non-master nodes/threads: */
840 if (!isSimulationMasterRank
)
842 inputrec
= &inputrecInstance
;
844 init_parallel(cr
, inputrec
, &mtop
, partialDeserializedTpr
.get());
846 GMX_RELEASE_ASSERT(inputrec
!= nullptr, "All ranks should have a valid inputrec now");
847 partialDeserializedTpr
.reset(nullptr);
849 // Now the number of ranks is known to all ranks, and each knows
850 // the inputrec read by the master rank. The ranks can now all run
851 // the task-deciding functions and will agree on the result
852 // without needing to communicate.
853 const bool useDomainDecomposition
= (PAR(cr
) && !(EI_TPI(inputrec
->eI
) || inputrec
->eI
== eiNM
));
855 // Note that these variables describe only their own node.
857 // Note that when bonded interactions run on a GPU they always run
858 // alongside a nonbonded task, so do not influence task assignment
859 // even though they affect the force calculation workload.
860 bool useGpuForNonbonded
= false;
861 bool useGpuForPme
= false;
862 bool useGpuForBonded
= false;
863 bool useGpuForUpdate
= false;
864 bool gpusWereDetected
= hwinfo
->ngpu_compatible_tot
> 0;
867 // It's possible that there are different numbers of GPUs on
868 // different nodes, which is the user's responsibility to
869 // handle. If unsuitable, we will notice that during task
871 auto canUseGpuForNonbonded
= buildSupportsNonbondedOnGpu(nullptr);
872 useGpuForNonbonded
= decideWhetherToUseGpusForNonbonded(
873 nonbondedTarget
, userGpuTaskAssignment
, emulateGpuNonbonded
, canUseGpuForNonbonded
,
874 gpuAccelerationOfNonbondedIsUseful(mdlog
, *inputrec
, !GMX_THREAD_MPI
), gpusWereDetected
);
875 useGpuForPme
= decideWhetherToUseGpusForPme(
876 useGpuForNonbonded
, pmeTarget
, userGpuTaskAssignment
, *hwinfo
, *inputrec
, mtop
,
877 cr
->nnodes
, domdecOptions
.numPmeRanks
, gpusWereDetected
);
878 auto canUseGpuForBonded
= buildSupportsGpuBondeds(nullptr)
879 && inputSupportsGpuBondeds(*inputrec
, mtop
, nullptr);
880 useGpuForBonded
= decideWhetherToUseGpusForBonded(
881 useGpuForNonbonded
, useGpuForPme
, bondedTarget
, canUseGpuForBonded
,
882 EVDW_PME(inputrec
->vdwtype
), EEL_PME_EWALD(inputrec
->coulombtype
),
883 domdecOptions
.numPmeRanks
, gpusWereDetected
);
885 pmeRunMode
= (useGpuForPme
? PmeRunMode::GPU
: PmeRunMode::CPU
);
886 if (pmeRunMode
== PmeRunMode::GPU
)
888 if (pmeFftTarget
== TaskTarget::Cpu
)
890 pmeRunMode
= PmeRunMode::Mixed
;
893 else if (pmeFftTarget
== TaskTarget::Gpu
)
896 "Assigning FFTs to GPU requires PME to be assigned to GPU as well. With PME "
897 "on CPU you should not be using -pmefft.");
900 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
902 // Initialize development feature flags that enabled by environment variable
903 // and report those features that are enabled.
904 const DevelopmentFeatureFlags devFlags
=
905 manageDevelopmentFeatures(mdlog
, useGpuForNonbonded
, pmeRunMode
);
908 // TODO: hide restraint implementation details from Mdrunner.
909 // There is nothing unique about restraints at this point as far as the
910 // Mdrunner is concerned. The Mdrunner should just be getting a sequence of
911 // factory functions from the SimulationContext on which to call mdModules_->add().
912 // TODO: capture all restraints into a single RestraintModule, passed to the runner builder.
913 for (auto&& restraint
: restraintManager_
->getRestraints())
915 auto module
= RestraintMDModule::create(restraint
, restraint
->sites());
916 mdModules_
->add(std::move(module
));
919 // TODO: Error handling
920 mdModules_
->assignOptionsToModules(*inputrec
->params
, nullptr);
921 const auto& mdModulesNotifier
= mdModules_
->notifier().notifier_
;
923 if (inputrec
->internalParameters
!= nullptr)
925 mdModulesNotifier
.notify(*inputrec
->internalParameters
);
928 if (fplog
!= nullptr)
930 pr_inputrec(fplog
, 0, "Input Parameters", inputrec
, FALSE
);
931 fprintf(fplog
, "\n");
936 /* In rerun, set velocities to zero if present */
937 if (doRerun
&& ((globalState
->flags
& (1 << estV
)) != 0))
939 // rerun does not use velocities
943 "Rerun trajectory contains velocities. Rerun does only evaluate "
944 "potential energy and forces. The velocities will be ignored.");
945 for (int i
= 0; i
< globalState
->natoms
; i
++)
947 clear_rvec(globalState
->v
[i
]);
949 globalState
->flags
&= ~(1 << estV
);
952 /* now make sure the state is initialized and propagated */
953 set_state_entries(globalState
.get(), inputrec
);
956 /* NM and TPI parallelize over force/energy calculations, not atoms,
957 * so we need to initialize and broadcast the global state.
959 if (inputrec
->eI
== eiNM
|| inputrec
->eI
== eiTPI
)
963 globalState
= std::make_unique
<t_state
>();
965 broadcastStateWithoutDynamics(cr
, globalState
.get());
968 /* A parallel command line option consistency check that we can
969 only do after any threads have started. */
971 && (domdecOptions
.numCells
[XX
] > 1 || domdecOptions
.numCells
[YY
] > 1
972 || domdecOptions
.numCells
[ZZ
] > 1 || domdecOptions
.numPmeRanks
> 0))
975 "The -dd or -npme option request a parallel simulation, "
977 "but %s was compiled without threads or MPI enabled",
978 output_env_get_program_display_name(oenv
));
980 "but the number of MPI-threads (option -ntmpi) is not set or is 1");
982 "but %s was not started through mpirun/mpiexec or only one rank was requested "
983 "through mpirun/mpiexec",
984 output_env_get_program_display_name(oenv
));
988 if (doRerun
&& (EI_ENERGY_MINIMIZATION(inputrec
->eI
) || eiNM
== inputrec
->eI
))
991 "The .mdp file specified an energy mininization or normal mode algorithm, and "
992 "these are not compatible with mdrun -rerun");
995 if (!(EEL_PME(inputrec
->coulombtype
) || EVDW_PME(inputrec
->vdwtype
)))
997 if (domdecOptions
.numPmeRanks
> 0)
999 gmx_fatal_collective(FARGS
, cr
->mpi_comm_mysim
, MASTER(cr
),
1000 "PME-only ranks are requested, but the system does not use PME "
1001 "for electrostatics or LJ");
1004 domdecOptions
.numPmeRanks
= 0;
1007 if (useGpuForNonbonded
&& domdecOptions
.numPmeRanks
< 0)
1009 /* With NB GPUs we don't automatically use PME-only CPU ranks. PME ranks can
1010 * improve performance with many threads per GPU, since our OpenMP
1011 * scaling is bad, but it's difficult to automate the setup.
1013 domdecOptions
.numPmeRanks
= 0;
1017 if (domdecOptions
.numPmeRanks
< 0)
1019 domdecOptions
.numPmeRanks
= 0;
1020 // TODO possibly print a note that one can opt-in for a separate PME GPU rank?
1024 GMX_RELEASE_ASSERT(domdecOptions
.numPmeRanks
<= 1,
1025 "PME GPU decomposition is not supported");
1032 fcRegisterSteps(inputrec
->nsteps
, inputrec
->init_step
);
1036 /* NMR restraints must be initialized before load_checkpoint,
1037 * since with time averaging the history is added to t_state.
1038 * For proper consistency check we therefore need to extend
1040 * So the PME-only nodes (if present) will also initialize
1041 * the distance restraints.
1045 /* This needs to be called before read_checkpoint to extend the state */
1046 init_disres(fplog
, &mtop
, inputrec
, cr
, ms
, fcd
, globalState
.get(), replExParams
.exchangeInterval
> 0);
1048 init_orires(fplog
, &mtop
, inputrec
, cr
, ms
, globalState
.get(), &(fcd
->orires
));
1050 auto deform
= prepareBoxDeformation(globalState
->box
, cr
, *inputrec
);
1052 ObservablesHistory observablesHistory
= {};
1054 if (startingBehavior
!= StartingBehavior::NewSimulation
)
1056 /* Check if checkpoint file exists before doing continuation.
1057 * This way we can use identical input options for the first and subsequent runs...
1059 if (mdrunOptions
.numStepsCommandline
> -2)
1061 /* Temporarily set the number of steps to unlimited to avoid
1062 * triggering the nsteps check in load_checkpoint().
1063 * This hack will go away soon when the -nsteps option is removed.
1065 inputrec
->nsteps
= -1;
1068 load_checkpoint(opt2fn_master("-cpi", filenames
.size(), filenames
.data(), cr
),
1069 logFileHandle
, cr
, domdecOptions
.numCells
, inputrec
, globalState
.get(),
1070 &observablesHistory
, mdrunOptions
.reproducible
, mdModules_
->notifier());
1072 if (startingBehavior
== StartingBehavior::RestartWithAppending
&& logFileHandle
)
1074 // Now we can start normal logging to the truncated log file.
1075 fplog
= gmx_fio_getfp(logFileHandle
);
1076 prepareLogAppending(fplog
);
1077 logOwner
= buildLogger(fplog
, MASTER(cr
));
1078 mdlog
= logOwner
.logger();
1082 if (mdrunOptions
.numStepsCommandline
> -2)
1087 "The -nsteps functionality is deprecated, and may be removed in a future "
1089 "Consider using gmx convert-tpr -nsteps or changing the appropriate .mdp "
1092 /* override nsteps with value set on the commandline */
1093 override_nsteps_cmdline(mdlog
, mdrunOptions
.numStepsCommandline
, inputrec
);
1097 copy_mat(globalState
->box
, box
);
1102 gmx_bcast(sizeof(box
), box
, cr
);
1105 if (inputrec
->cutoff_scheme
!= ecutsVERLET
)
1108 "This group-scheme .tpr file can no longer be run by mdrun. Please update to the "
1109 "Verlet scheme, or use an earlier version of GROMACS if necessary.");
1111 /* Update rlist and nstlist. */
1112 prepare_verlet_scheme(fplog
, cr
, inputrec
, nstlist_cmdline
, &mtop
, box
,
1113 useGpuForNonbonded
|| (emulateGpuNonbonded
== EmulateGpuNonbonded::Yes
),
1116 const bool prefer1DAnd1PulseDD
= (devFlags
.enableGpuHaloExchange
&& useGpuForNonbonded
);
1117 // This builder is necessary while we have multi-part construction
1118 // of DD. Before DD is constructed, we use the existence of
1119 // the builder object to indicate that further construction of DD
1121 std::unique_ptr
<DomainDecompositionBuilder
> ddBuilder
;
1122 if (useDomainDecomposition
)
1124 ddBuilder
= std::make_unique
<DomainDecompositionBuilder
>(
1125 mdlog
, cr
, domdecOptions
, mdrunOptions
, prefer1DAnd1PulseDD
, mtop
, *inputrec
, box
,
1126 positionsFromStatePointer(globalState
.get()));
1130 /* PME, if used, is done on all nodes with 1D decomposition */
1132 cr
->duty
= (DUTY_PP
| DUTY_PME
);
1134 if (inputrec
->pbcType
== PbcType::Screw
)
1136 gmx_fatal(FARGS
, "pbc=screw is only implemented with domain decomposition");
1140 // Produce the task assignment for this rank.
1141 GpuTaskAssignmentsBuilder gpuTaskAssignmentsBuilder
;
1142 GpuTaskAssignments gpuTaskAssignments
= gpuTaskAssignmentsBuilder
.build(
1143 gpuIdsToUse
, userGpuTaskAssignment
, *hwinfo
, cr
, ms
, physicalNodeComm
, nonbondedTarget
,
1144 pmeTarget
, bondedTarget
, updateTarget
, useGpuForNonbonded
, useGpuForPme
,
1145 thisRankHasDuty(cr
, DUTY_PP
),
1146 // TODO cr->duty & DUTY_PME should imply that a PME
1147 // algorithm is active, but currently does not.
1148 EEL_PME(inputrec
->coulombtype
) && thisRankHasDuty(cr
, DUTY_PME
));
1150 const bool printHostName
= (cr
->nnodes
> 1);
1151 gpuTaskAssignments
.reportGpuUsage(mdlog
, printHostName
, useGpuForBonded
, pmeRunMode
);
1153 // If the user chose a task assignment, give them some hints
1154 // where appropriate.
1155 if (!userGpuTaskAssignment
.empty())
1157 gpuTaskAssignments
.logPerformanceHints(mdlog
, ssize(gpuIdsToUse
));
1160 // Get the device handles for the modules, nullptr when no task is assigned.
1161 gmx_device_info_t
* nonbondedDeviceInfo
= gpuTaskAssignments
.initNonbondedDevice(cr
);
1162 gmx_device_info_t
* pmeDeviceInfo
= gpuTaskAssignments
.initPmeDevice();
1164 // TODO Initialize GPU streams here.
1166 // TODO Currently this is always built, yet DD partition code
1167 // checks if it is built before using it. Probably it should
1168 // become an MDModule that is made only when another module
1169 // requires it (e.g. pull, CompEl, density fitting), so that we
1170 // don't update the local atom sets unilaterally every step.
1171 LocalAtomSetManager atomSets
;
1174 // TODO Pass the GPU streams to ddBuilder to use in buffer
1175 // transfers (e.g. halo exchange)
1176 cr
->dd
= ddBuilder
->build(&atomSets
);
1177 // The builder's job is done, so destruct it
1178 ddBuilder
.reset(nullptr);
1179 // Note that local state still does not exist yet.
1182 // The GPU update is decided here because we need to know whether the constraints or
1183 // SETTLEs can span accross the domain borders (i.e. whether or not update groups are
1184 // defined). This is only known after DD is initialized, hence decision on using GPU
1185 // update is done so late.
1188 const bool useUpdateGroups
= cr
->dd
? ddUsesUpdateGroups(*cr
->dd
) : false;
1190 useGpuForUpdate
= decideWhetherToUseGpuForUpdate(
1191 devFlags
.forceGpuUpdateDefault
, useDomainDecomposition
, useUpdateGroups
, pmeRunMode
,
1192 domdecOptions
.numPmeRanks
> 0, useGpuForNonbonded
, updateTarget
, gpusWereDetected
,
1193 *inputrec
, mtop
, doEssentialDynamics
, gmx_mtop_ftype_count(mtop
, F_ORIRES
) > 0,
1194 replExParams
.exchangeInterval
> 0, doRerun
, mdlog
);
1196 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1200 /* After possible communicator splitting in make_dd_communicators.
1201 * we can set up the intra/inter node communication.
1203 gmx_setup_nodecomm(fplog
, cr
);
1209 GMX_LOG(mdlog
.warning
)
1211 .appendTextFormatted(
1212 "This is simulation %d out of %d running as a composite GROMACS\n"
1213 "multi-simulation job. Setup for this simulation:\n",
1216 GMX_LOG(mdlog
.warning
)
1217 .appendTextFormatted("Using %d MPI %s\n", cr
->nnodes
,
1219 cr
->nnodes
== 1 ? "thread" : "threads"
1221 cr
->nnodes
== 1 ? "process" : "processes"
1227 // If mdrun -pin auto honors any affinity setting that already
1228 // exists. If so, it is nice to provide feedback about whether
1229 // that existing affinity setting was from OpenMP or something
1230 // else, so we run this code both before and after we initialize
1231 // the OpenMP support.
1232 gmx_check_thread_affinity_set(mdlog
, &hw_opt
, hwinfo
->nthreads_hw_avail
, FALSE
);
1233 /* Check and update the number of OpenMP threads requested */
1234 checkAndUpdateRequestedNumOpenmpThreads(&hw_opt
, *hwinfo
, cr
, ms
, physicalNodeComm
.size_
,
1235 pmeRunMode
, mtop
, *inputrec
);
1237 gmx_omp_nthreads_init(mdlog
, cr
, hwinfo
->nthreads_hw_avail
, physicalNodeComm
.size_
,
1238 hw_opt
.nthreads_omp
, hw_opt
.nthreads_omp_pme
, !thisRankHasDuty(cr
, DUTY_PP
));
1240 // Enable FP exception detection, but not in
1241 // Release mode and not for compilers with known buggy FP
1242 // exception support (clang with any optimization) or suspected
1243 // buggy FP exception support (gcc 7.* with optimization).
1244 #if !defined NDEBUG \
1245 && !((defined __clang__ || (defined(__GNUC__) && !defined(__ICC) && __GNUC__ == 7)) \
1246 && defined __OPTIMIZE__)
1247 const bool bEnableFPE
= true;
1249 const bool bEnableFPE
= false;
1251 // FIXME - reconcile with gmx_feenableexcept() call from CommandLineModuleManager::run()
1254 gmx_feenableexcept();
1257 /* Now that we know the setup is consistent, check for efficiency */
1258 check_resource_division_efficiency(hwinfo
, gpuTaskAssignments
.thisRankHasAnyGpuTask(),
1259 mdrunOptions
.ntompOptionIsSet
, cr
, mdlog
);
1261 /* getting number of PP/PME threads on this MPI / tMPI rank.
1262 PME: env variable should be read only on one node to make sure it is
1263 identical everywhere;
1265 const int numThreadsOnThisRank
= thisRankHasDuty(cr
, DUTY_PP
) ? gmx_omp_nthreads_get(emntNonbonded
)
1266 : gmx_omp_nthreads_get(emntPME
);
1267 checkHardwareOversubscription(numThreadsOnThisRank
, cr
->nodeid
, *hwinfo
->hardwareTopology
,
1268 physicalNodeComm
, mdlog
);
1270 // Enable Peer access between GPUs where available
1271 // Only for DD, only master PP rank needs to perform setup, and only if thread MPI plus
1272 // any of the GPU communication features are active.
1273 if (DOMAINDECOMP(cr
) && MASTER(cr
) && thisRankHasDuty(cr
, DUTY_PP
) && GMX_THREAD_MPI
1274 && (devFlags
.enableGpuHaloExchange
|| devFlags
.enableGpuPmePPComm
))
1276 setupGpuDevicePeerAccess(gpuIdsToUse
, mdlog
);
1279 if (hw_opt
.threadAffinity
!= ThreadAffinity::Off
)
1281 /* Before setting affinity, check whether the affinity has changed
1282 * - which indicates that probably the OpenMP library has changed it
1283 * since we first checked).
1285 gmx_check_thread_affinity_set(mdlog
, &hw_opt
, hwinfo
->nthreads_hw_avail
, TRUE
);
1287 int numThreadsOnThisNode
, intraNodeThreadOffset
;
1288 analyzeThreadsOnThisNode(physicalNodeComm
, numThreadsOnThisRank
, &numThreadsOnThisNode
,
1289 &intraNodeThreadOffset
);
1291 /* Set the CPU affinity */
1292 gmx_set_thread_affinity(mdlog
, cr
, &hw_opt
, *hwinfo
->hardwareTopology
, numThreadsOnThisRank
,
1293 numThreadsOnThisNode
, intraNodeThreadOffset
, nullptr);
1296 if (mdrunOptions
.timingOptions
.resetStep
> -1)
1301 "The -resetstep functionality is deprecated, and may be removed in a "
1304 wcycle
= wallcycle_init(fplog
, mdrunOptions
.timingOptions
.resetStep
, cr
);
1308 /* Master synchronizes its value of reset_counters with all nodes
1309 * including PME only nodes */
1310 int64_t reset_counters
= wcycle_get_reset_counters(wcycle
);
1311 gmx_bcast_sim(sizeof(reset_counters
), &reset_counters
, cr
);
1312 wcycle_set_reset_counters(wcycle
, reset_counters
);
1315 // Membrane embedding must be initialized before we call init_forcerec()
1320 fprintf(stderr
, "Initializing membed");
1322 /* Note that membed cannot work in parallel because mtop is
1323 * changed here. Fix this if we ever want to make it run with
1324 * multiple ranks. */
1325 membed
= init_membed(fplog
, filenames
.size(), filenames
.data(), &mtop
, inputrec
,
1326 globalState
.get(), cr
, &mdrunOptions
.checkpointOptions
.period
);
1329 const bool thisRankHasPmeGpuTask
= gpuTaskAssignments
.thisRankHasPmeGpuTask();
1330 std::unique_ptr
<MDAtoms
> mdAtoms
;
1331 std::unique_ptr
<gmx_vsite_t
> vsite
;
1334 if (thisRankHasDuty(cr
, DUTY_PP
))
1336 mdModulesNotifier
.notify(*cr
);
1337 mdModulesNotifier
.notify(&atomSets
);
1338 mdModulesNotifier
.notify(PeriodicBoundaryConditionType
{ inputrec
->pbcType
});
1339 mdModulesNotifier
.notify(SimulationTimeStep
{ inputrec
->delta_t
});
1340 /* Initiate forcerecord */
1341 fr
= new t_forcerec
;
1342 fr
->forceProviders
= mdModules_
->initForceProviders();
1343 init_forcerec(fplog
, mdlog
, fr
, fcd
, inputrec
, &mtop
, cr
, box
,
1344 opt2fn("-table", filenames
.size(), filenames
.data()),
1345 opt2fn("-tablep", filenames
.size(), filenames
.data()),
1346 opt2fns("-tableb", filenames
.size(), filenames
.data()), *hwinfo
,
1347 nonbondedDeviceInfo
, useGpuForBonded
,
1348 pmeRunMode
== PmeRunMode::GPU
&& !thisRankHasDuty(cr
, DUTY_PME
), pforce
, wcycle
);
1350 // TODO Move this to happen during domain decomposition setup,
1351 // once stream and event handling works well with that.
1352 // TODO remove need to pass local stream into GPU halo exchange - Redmine #3093
1353 if (havePPDomainDecomposition(cr
) && prefer1DAnd1PulseDD
&& is1DAnd1PulseDD(*cr
->dd
))
1355 GMX_RELEASE_ASSERT(devFlags
.enableGpuBufferOps
,
1356 "Must use GMX_USE_GPU_BUFFER_OPS=1 to use GMX_GPU_DD_COMMS=1");
1358 Nbnxm::gpu_get_command_stream(fr
->nbv
->gpu_nbv
, InteractionLocality::Local
);
1359 void* streamNonLocal
=
1360 Nbnxm::gpu_get_command_stream(fr
->nbv
->gpu_nbv
, InteractionLocality::NonLocal
);
1361 GMX_LOG(mdlog
.warning
)
1363 .appendTextFormatted(
1364 "NOTE: This run uses the 'GPU halo exchange' feature, enabled by the "
1365 "GMX_GPU_DD_COMMS environment variable.");
1366 cr
->dd
->gpuHaloExchange
= std::make_unique
<GpuHaloExchange
>(
1367 cr
->dd
, cr
->mpi_comm_mysim
, streamLocal
, streamNonLocal
);
1370 /* Initialize the mdAtoms structure.
1371 * mdAtoms is not filled with atom data,
1372 * as this can not be done now with domain decomposition.
1374 mdAtoms
= makeMDAtoms(fplog
, mtop
, *inputrec
, thisRankHasPmeGpuTask
);
1375 if (globalState
&& thisRankHasPmeGpuTask
)
1377 // The pinning of coordinates in the global state object works, because we only use
1378 // PME on GPU without DD or on a separate PME rank, and because the local state pointer
1379 // points to the global state object without DD.
1380 // FIXME: MD and EM separately set up the local state - this should happen in the same
1381 // function, which should also perform the pinning.
1382 changePinningPolicy(&globalState
->x
, pme_get_pinning_policy());
1385 /* Initialize the virtual site communication */
1386 vsite
= initVsite(mtop
, cr
);
1388 calc_shifts(box
, fr
->shift_vec
);
1390 /* With periodic molecules the charge groups should be whole at start up
1391 * and the virtual sites should not be far from their proper positions.
1393 if (!inputrec
->bContinuation
&& MASTER(cr
)
1394 && !(inputrec
->pbcType
!= PbcType::No
&& inputrec
->bPeriodicMols
))
1396 /* Make molecules whole at start of run */
1397 if (fr
->pbcType
!= PbcType::No
)
1399 do_pbc_first_mtop(fplog
, inputrec
->pbcType
, box
, &mtop
, globalState
->x
.rvec_array());
1403 /* Correct initial vsite positions are required
1404 * for the initial distribution in the domain decomposition
1405 * and for the initial shell prediction.
1407 constructVsitesGlobal(mtop
, globalState
->x
);
1411 if (EEL_PME(fr
->ic
->eeltype
) || EVDW_PME(fr
->ic
->vdwtype
))
1413 ewaldcoeff_q
= fr
->ic
->ewaldcoeff_q
;
1414 ewaldcoeff_lj
= fr
->ic
->ewaldcoeff_lj
;
1419 /* This is a PME only node */
1421 GMX_ASSERT(globalState
== nullptr,
1422 "We don't need the state on a PME only rank and expect it to be unitialized");
1424 ewaldcoeff_q
= calc_ewaldcoeff_q(inputrec
->rcoulomb
, inputrec
->ewald_rtol
);
1425 ewaldcoeff_lj
= calc_ewaldcoeff_lj(inputrec
->rvdw
, inputrec
->ewald_rtol_lj
);
1428 gmx_pme_t
* sepPmeData
= nullptr;
1429 // This reference hides the fact that PME data is owned by runner on PME-only ranks and by forcerec on other ranks
1430 GMX_ASSERT(thisRankHasDuty(cr
, DUTY_PP
) == (fr
!= nullptr),
1431 "Double-checking that only PME-only ranks have no forcerec");
1432 gmx_pme_t
*& pmedata
= fr
? fr
->pmedata
: sepPmeData
;
1434 // TODO should live in ewald module once its testing is improved
1436 // Later, this program could contain kernels that might be later
1437 // re-used as auto-tuning progresses, or subsequent simulations
1439 PmeGpuProgramStorage pmeGpuProgram
;
1440 if (thisRankHasPmeGpuTask
)
1442 pmeGpuProgram
= buildPmeGpuProgram(pmeDeviceInfo
);
1445 /* Initiate PME if necessary,
1446 * either on all nodes or on dedicated PME nodes only. */
1447 if (EEL_PME(inputrec
->coulombtype
) || EVDW_PME(inputrec
->vdwtype
))
1449 if (mdAtoms
&& mdAtoms
->mdatoms())
1451 nChargePerturbed
= mdAtoms
->mdatoms()->nChargePerturbed
;
1452 if (EVDW_PME(inputrec
->vdwtype
))
1454 nTypePerturbed
= mdAtoms
->mdatoms()->nTypePerturbed
;
1457 if (cr
->npmenodes
> 0)
1459 /* The PME only nodes need to know nChargePerturbed(FEP on Q) and nTypePerturbed(FEP on LJ)*/
1460 gmx_bcast_sim(sizeof(nChargePerturbed
), &nChargePerturbed
, cr
);
1461 gmx_bcast_sim(sizeof(nTypePerturbed
), &nTypePerturbed
, cr
);
1464 if (thisRankHasDuty(cr
, DUTY_PME
))
1468 pmedata
= gmx_pme_init(cr
, getNumPmeDomains(cr
->dd
), inputrec
, nChargePerturbed
!= 0,
1469 nTypePerturbed
!= 0, mdrunOptions
.reproducible
, ewaldcoeff_q
,
1470 ewaldcoeff_lj
, gmx_omp_nthreads_get(emntPME
), pmeRunMode
,
1471 nullptr, pmeDeviceInfo
, pmeGpuProgram
.get(), mdlog
);
1473 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1478 if (EI_DYNAMICS(inputrec
->eI
))
1480 /* Turn on signal handling on all nodes */
1482 * (A user signal from the PME nodes (if any)
1483 * is communicated to the PP nodes.
1485 signal_handler_install();
1488 pull_t
* pull_work
= nullptr;
1489 if (thisRankHasDuty(cr
, DUTY_PP
))
1491 /* Assumes uniform use of the number of OpenMP threads */
1492 walltime_accounting
= walltime_accounting_init(gmx_omp_nthreads_get(emntDefault
));
1494 if (inputrec
->bPull
)
1496 /* Initialize pull code */
1497 pull_work
= init_pull(fplog
, inputrec
->pull
, inputrec
, &mtop
, cr
, &atomSets
,
1498 inputrec
->fepvals
->init_lambda
);
1499 if (inputrec
->pull
->bXOutAverage
|| inputrec
->pull
->bFOutAverage
)
1501 initPullHistory(pull_work
, &observablesHistory
);
1503 if (EI_DYNAMICS(inputrec
->eI
) && MASTER(cr
))
1505 init_pull_output_files(pull_work
, filenames
.size(), filenames
.data(), oenv
, startingBehavior
);
1509 std::unique_ptr
<EnforcedRotation
> enforcedRotation
;
1512 /* Initialize enforced rotation code */
1514 init_rot(fplog
, inputrec
, filenames
.size(), filenames
.data(), cr
, &atomSets
,
1515 globalState
.get(), &mtop
, oenv
, mdrunOptions
, startingBehavior
);
1518 t_swap
* swap
= nullptr;
1519 if (inputrec
->eSwapCoords
!= eswapNO
)
1521 /* Initialize ion swapping code */
1522 swap
= init_swapcoords(fplog
, inputrec
,
1523 opt2fn_master("-swap", filenames
.size(), filenames
.data(), cr
),
1524 &mtop
, globalState
.get(), &observablesHistory
, cr
, &atomSets
,
1525 oenv
, mdrunOptions
, startingBehavior
);
1528 /* Let makeConstraints know whether we have essential dynamics constraints. */
1529 auto constr
= makeConstraints(mtop
, *inputrec
, pull_work
, doEssentialDynamics
, fplog
,
1530 *mdAtoms
->mdatoms(), cr
, ms
, &nrnb
, wcycle
, fr
->bMolPBC
);
1532 /* Energy terms and groups */
1533 gmx_enerdata_t
enerd(mtop
.groups
.groups
[SimulationAtomGroupType::EnergyOutput
].size(),
1534 inputrec
->fepvals
->n_lambda
);
1536 /* Kinetic energy data */
1537 gmx_ekindata_t ekind
;
1538 init_ekindata(fplog
, &mtop
, &(inputrec
->opts
), &ekind
, inputrec
->cos_accel
);
1540 /* Set up interactive MD (IMD) */
1542 makeImdSession(inputrec
, cr
, wcycle
, &enerd
, ms
, &mtop
, mdlog
,
1543 MASTER(cr
) ? globalState
->x
.rvec_array() : nullptr, filenames
.size(),
1544 filenames
.data(), oenv
, mdrunOptions
.imdOptions
, startingBehavior
);
1546 if (DOMAINDECOMP(cr
))
1548 GMX_RELEASE_ASSERT(fr
, "fr was NULL while cr->duty was DUTY_PP");
1549 /* This call is not included in init_domain_decomposition mainly
1550 * because fr->cginfo_mb is set later.
1552 dd_init_bondeds(fplog
, cr
->dd
, mtop
, vsite
.get(), inputrec
,
1553 domdecOptions
.checkBondedInteractions
, fr
->cginfo_mb
);
1556 const bool inputIsCompatibleWithModularSimulator
= ModularSimulator::isInputCompatible(
1557 false, inputrec
, doRerun
, vsite
.get(), ms
, replExParams
, fcd
,
1558 static_cast<int>(filenames
.size()), filenames
.data(), &observablesHistory
, membed
);
1560 const bool useModularSimulator
= inputIsCompatibleWithModularSimulator
1561 && !(getenv("GMX_DISABLE_MODULAR_SIMULATOR") != nullptr);
1563 // TODO This is not the right place to manage the lifetime of
1564 // this data structure, but currently it's the easiest way to
1566 MdrunScheduleWorkload runScheduleWork
;
1567 // Also populates the simulation constant workload description.
1568 runScheduleWork
.simulationWork
= createSimulationWorkload(
1569 useGpuForNonbonded
, pmeRunMode
, useGpuForBonded
, useGpuForUpdate
,
1570 devFlags
.enableGpuBufferOps
, devFlags
.enableGpuHaloExchange
,
1571 devFlags
.enableGpuPmePPComm
, haveEwaldSurfaceContribution(*inputrec
));
1573 std::unique_ptr
<gmx::StatePropagatorDataGpu
> stateGpu
;
1574 if (gpusWereDetected
1575 && ((useGpuForPme
&& thisRankHasDuty(cr
, DUTY_PME
))
1576 || runScheduleWork
.simulationWork
.useGpuBufferOps
))
1578 const void* pmeStream
= pme_gpu_get_device_stream(fr
->pmedata
);
1579 const void* localStream
=
1580 fr
->nbv
->gpu_nbv
!= nullptr
1581 ? Nbnxm::gpu_get_command_stream(fr
->nbv
->gpu_nbv
, InteractionLocality::Local
)
1583 const void* nonLocalStream
=
1584 fr
->nbv
->gpu_nbv
!= nullptr
1585 ? Nbnxm::gpu_get_command_stream(fr
->nbv
->gpu_nbv
, InteractionLocality::NonLocal
)
1587 const void* deviceContext
= pme_gpu_get_device_context(fr
->pmedata
);
1588 const int paddingSize
= pme_gpu_get_padding_size(fr
->pmedata
);
1589 GpuApiCallBehavior transferKind
= (inputrec
->eI
== eiMD
&& !doRerun
&& !useModularSimulator
)
1590 ? GpuApiCallBehavior::Async
1591 : GpuApiCallBehavior::Sync
;
1593 stateGpu
= std::make_unique
<gmx::StatePropagatorDataGpu
>(
1594 pmeStream
, localStream
, nonLocalStream
, deviceContext
, transferKind
, paddingSize
, wcycle
);
1595 fr
->stateGpu
= stateGpu
.get();
1598 GMX_ASSERT(stopHandlerBuilder_
, "Runner must provide StopHandlerBuilder to simulator.");
1599 SimulatorBuilder simulatorBuilder
;
1601 // build and run simulator object based on user-input
1602 auto simulator
= simulatorBuilder
.build(
1603 inputIsCompatibleWithModularSimulator
, fplog
, cr
, ms
, mdlog
,
1604 static_cast<int>(filenames
.size()), filenames
.data(), oenv
, mdrunOptions
,
1605 startingBehavior
, vsite
.get(), constr
.get(),
1606 enforcedRotation
? enforcedRotation
->getLegacyEnfrot() : nullptr, deform
.get(),
1607 mdModules_
->outputProvider(), mdModules_
->notifier(), inputrec
, imdSession
.get(),
1608 pull_work
, swap
, &mtop
, fcd
, globalState
.get(), &observablesHistory
, mdAtoms
.get(),
1609 &nrnb
, wcycle
, fr
, &enerd
, &ekind
, &runScheduleWork
, replExParams
, membed
,
1610 walltime_accounting
, std::move(stopHandlerBuilder_
), doRerun
);
1613 if (fr
->pmePpCommGpu
)
1615 // destroy object since it is no longer required. (This needs to be done while the GPU context still exists.)
1616 fr
->pmePpCommGpu
.reset();
1619 if (inputrec
->bPull
)
1621 finish_pull(pull_work
);
1623 finish_swapcoords(swap
);
1627 GMX_RELEASE_ASSERT(pmedata
, "pmedata was NULL while cr->duty was not DUTY_PP");
1629 walltime_accounting
= walltime_accounting_init(gmx_omp_nthreads_get(emntPME
));
1630 gmx_pmeonly(pmedata
, cr
, &nrnb
, wcycle
, walltime_accounting
, inputrec
, pmeRunMode
);
1633 wallcycle_stop(wcycle
, ewcRUN
);
1635 /* Finish up, write some stuff
1636 * if rerunMD, don't write last frame again
1638 finish_run(fplog
, mdlog
, cr
, inputrec
, &nrnb
, wcycle
, walltime_accounting
,
1639 fr
? fr
->nbv
.get() : nullptr, pmedata
, EI_DYNAMICS(inputrec
->eI
) && !isMultiSim(ms
));
1641 // clean up cycle counter
1642 wallcycle_destroy(wcycle
);
1647 gmx_pme_destroy(pmedata
);
1651 // FIXME: this is only here to manually unpin mdAtoms->chargeA_ and state->x,
1652 // before we destroy the GPU context(s) in free_gpu().
1653 // Pinned buffers are associated with contexts in CUDA.
1654 // As soon as we destroy GPU contexts after mdrunner() exits, these lines should go.
1655 mdAtoms
.reset(nullptr);
1656 globalState
.reset(nullptr);
1657 mdModules_
.reset(nullptr); // destruct force providers here as they might also use the GPU
1658 /* Free pinned buffers in *fr */
1662 if (hwinfo
->gpu_info
.n_dev
> 0)
1664 /* stop the GPU profiler (only CUDA) */
1668 /* With tMPI we need to wait for all ranks to finish deallocation before
1669 * destroying the CUDA context in free_gpu() as some tMPI ranks may be sharing
1672 * This is not a concern in OpenCL where we use one context per rank which
1673 * is freed in nbnxn_gpu_free().
1675 * Note: it is safe to not call the barrier on the ranks which do not use GPU,
1676 * but it is easier and more futureproof to call it on the whole node.
1678 * Note that this function needs to be called even if GPUs are not used
1679 * in this run because the PME ranks have no knowledge of whether GPUs
1680 * are used or not, but all ranks need to enter the barrier below.
1681 * \todo Remove this physical node barrier after making sure
1682 * that it's not needed anymore (with a shared GPU run).
1686 physicalNodeComm
.barrier();
1689 free_gpu(nonbondedDeviceInfo
);
1690 free_gpu(pmeDeviceInfo
);
1695 free_membed(membed
);
1698 /* Does what it says */
1699 print_date_and_time(fplog
, cr
->nodeid
, "Finished mdrun", gmx_gettime());
1700 walltime_accounting_destroy(walltime_accounting
);
1702 // Ensure log file content is written
1705 gmx_fio_flush(logFileHandle
);
1708 /* Reset FPEs (important for unit tests) by disabling them. Assumes no
1709 * exceptions were enabled before function was called. */
1712 gmx_fedisableexcept();
1715 auto rc
= static_cast<int>(gmx_get_stop_condition());
1718 /* we need to join all threads. The sub-threads join when they
1719 exit this function, but the master thread needs to be told to
1721 if (PAR(cr
) && MASTER(cr
))
1729 Mdrunner::~Mdrunner()
1731 // Clean up of the Manager.
1732 // This will end up getting called on every thread-MPI rank, which is unnecessary,
1733 // but okay as long as threads synchronize some time before adding or accessing
1734 // a new set of restraints.
1735 if (restraintManager_
)
1737 restraintManager_
->clear();
1738 GMX_ASSERT(restraintManager_
->countRestraints() == 0,
1739 "restraints added during runner life time should be cleared at runner "
1744 void Mdrunner::addPotential(std::shared_ptr
<gmx::IRestraintPotential
> puller
, const std::string
& name
)
1746 GMX_ASSERT(restraintManager_
, "Mdrunner must have a restraint manager.");
1747 // Not sure if this should be logged through the md logger or something else,
1748 // but it is helpful to have some sort of INFO level message sent somewhere.
1749 // std::cout << "Registering restraint named " << name << std::endl;
1751 // When multiple restraints are used, it may be wasteful to register them separately.
1752 // Maybe instead register an entire Restraint Manager as a force provider.
1753 restraintManager_
->addToSpec(std::move(puller
), name
);
1756 Mdrunner::Mdrunner(std::unique_ptr
<MDModules
> mdModules
) : mdModules_(std::move(mdModules
)) {}
1758 Mdrunner::Mdrunner(Mdrunner
&&) noexcept
= default;
1760 //NOLINTNEXTLINE(performance-noexcept-move-constructor) working around GCC bug 58265
1761 Mdrunner
& Mdrunner::operator=(Mdrunner
&& /*handle*/) noexcept(BUGFREE_NOEXCEPT_STRING
) = default;
1763 class Mdrunner::BuilderImplementation
1766 BuilderImplementation() = delete;
1767 BuilderImplementation(std::unique_ptr
<MDModules
> mdModules
, compat::not_null
<SimulationContext
*> context
);
1768 ~BuilderImplementation();
1770 BuilderImplementation
& setExtraMdrunOptions(const MdrunOptions
& options
,
1771 real forceWarningThreshold
,
1772 StartingBehavior startingBehavior
);
1774 void addDomdec(const DomdecOptions
& options
);
1776 void addVerletList(int nstlist
);
1778 void addReplicaExchange(const ReplicaExchangeParameters
& params
);
1780 void addNonBonded(const char* nbpu_opt
);
1782 void addPME(const char* pme_opt_
, const char* pme_fft_opt_
);
1784 void addBondedTaskAssignment(const char* bonded_opt
);
1786 void addUpdateTaskAssignment(const char* update_opt
);
1788 void addHardwareOptions(const gmx_hw_opt_t
& hardwareOptions
);
1790 void addFilenames(ArrayRef
<const t_filenm
> filenames
);
1792 void addOutputEnvironment(gmx_output_env_t
* outputEnvironment
);
1794 void addLogFile(t_fileio
* logFileHandle
);
1796 void addStopHandlerBuilder(std::unique_ptr
<StopHandlerBuilder
> builder
);
1801 // Default parameters copied from runner.h
1802 // \todo Clarify source(s) of default parameters.
1804 const char* nbpu_opt_
= nullptr;
1805 const char* pme_opt_
= nullptr;
1806 const char* pme_fft_opt_
= nullptr;
1807 const char* bonded_opt_
= nullptr;
1808 const char* update_opt_
= nullptr;
1810 MdrunOptions mdrunOptions_
;
1812 DomdecOptions domdecOptions_
;
1814 ReplicaExchangeParameters replicaExchangeParameters_
;
1816 //! Command-line override for the duration of a neighbor list with the Verlet scheme.
1819 //! Multisim communicator handle.
1820 gmx_multisim_t
* multiSimulation_
;
1822 //! mdrun communicator
1823 MPI_Comm communicator_
= MPI_COMM_NULL
;
1825 //! Print a warning if any force is larger than this (in kJ/mol nm).
1826 real forceWarningThreshold_
= -1;
1828 //! Whether the simulation will start afresh, or restart with/without appending.
1829 StartingBehavior startingBehavior_
= StartingBehavior::NewSimulation
;
1831 //! The modules that comprise the functionality of mdrun.
1832 std::unique_ptr
<MDModules
> mdModules_
;
1834 //! \brief Parallelism information.
1835 gmx_hw_opt_t hardwareOptions_
;
1837 //! filename options for simulation.
1838 ArrayRef
<const t_filenm
> filenames_
;
1840 /*! \brief Handle to output environment.
1842 * \todo gmx_output_env_t needs lifetime management.
1844 gmx_output_env_t
* outputEnvironment_
= nullptr;
1846 /*! \brief Non-owning handle to MD log file.
1848 * \todo Context should own output facilities for client.
1849 * \todo Improve log file handle management.
1851 * Code managing the FILE* relies on the ability to set it to
1852 * nullptr to check whether the filehandle is valid.
1854 t_fileio
* logFileHandle_
= nullptr;
1857 * \brief Builder for simulation stop signal handler.
1859 std::unique_ptr
<StopHandlerBuilder
> stopHandlerBuilder_
= nullptr;
1862 Mdrunner::BuilderImplementation::BuilderImplementation(std::unique_ptr
<MDModules
> mdModules
,
1863 compat::not_null
<SimulationContext
*> context
) :
1864 mdModules_(std::move(mdModules
))
1866 communicator_
= context
->communicator_
;
1867 multiSimulation_
= context
->multiSimulation_
.get();
1870 Mdrunner::BuilderImplementation::~BuilderImplementation() = default;
1872 Mdrunner::BuilderImplementation
&
1873 Mdrunner::BuilderImplementation::setExtraMdrunOptions(const MdrunOptions
& options
,
1874 const real forceWarningThreshold
,
1875 const StartingBehavior startingBehavior
)
1877 mdrunOptions_
= options
;
1878 forceWarningThreshold_
= forceWarningThreshold
;
1879 startingBehavior_
= startingBehavior
;
1883 void Mdrunner::BuilderImplementation::addDomdec(const DomdecOptions
& options
)
1885 domdecOptions_
= options
;
1888 void Mdrunner::BuilderImplementation::addVerletList(int nstlist
)
1893 void Mdrunner::BuilderImplementation::addReplicaExchange(const ReplicaExchangeParameters
& params
)
1895 replicaExchangeParameters_
= params
;
1898 Mdrunner
Mdrunner::BuilderImplementation::build()
1900 auto newRunner
= Mdrunner(std::move(mdModules_
));
1902 newRunner
.mdrunOptions
= mdrunOptions_
;
1903 newRunner
.pforce
= forceWarningThreshold_
;
1904 newRunner
.startingBehavior
= startingBehavior_
;
1905 newRunner
.domdecOptions
= domdecOptions_
;
1907 // \todo determine an invariant to check or confirm that all gmx_hw_opt_t objects are valid
1908 newRunner
.hw_opt
= hardwareOptions_
;
1910 // No invariant to check. This parameter exists to optionally override other behavior.
1911 newRunner
.nstlist_cmdline
= nstlist_
;
1913 newRunner
.replExParams
= replicaExchangeParameters_
;
1915 newRunner
.filenames
= filenames_
;
1917 newRunner
.communicator
= communicator_
;
1919 // nullptr is a valid value for the multisim handle
1920 newRunner
.ms
= multiSimulation_
;
1922 // \todo Clarify ownership and lifetime management for gmx_output_env_t
1923 // \todo Update sanity checking when output environment has clearly specified invariants.
1924 // Initialization and default values for oenv are not well specified in the current version.
1925 if (outputEnvironment_
)
1927 newRunner
.oenv
= outputEnvironment_
;
1931 GMX_THROW(gmx::APIError(
1932 "MdrunnerBuilder::addOutputEnvironment() is required before build()"));
1935 newRunner
.logFileHandle
= logFileHandle_
;
1939 newRunner
.nbpu_opt
= nbpu_opt_
;
1943 GMX_THROW(gmx::APIError("MdrunnerBuilder::addNonBonded() is required before build()"));
1946 if (pme_opt_
&& pme_fft_opt_
)
1948 newRunner
.pme_opt
= pme_opt_
;
1949 newRunner
.pme_fft_opt
= pme_fft_opt_
;
1953 GMX_THROW(gmx::APIError("MdrunnerBuilder::addElectrostatics() is required before build()"));
1958 newRunner
.bonded_opt
= bonded_opt_
;
1962 GMX_THROW(gmx::APIError(
1963 "MdrunnerBuilder::addBondedTaskAssignment() is required before build()"));
1968 newRunner
.update_opt
= update_opt_
;
1972 GMX_THROW(gmx::APIError(
1973 "MdrunnerBuilder::addUpdateTaskAssignment() is required before build() "));
1977 newRunner
.restraintManager_
= std::make_unique
<gmx::RestraintManager
>();
1979 if (stopHandlerBuilder_
)
1981 newRunner
.stopHandlerBuilder_
= std::move(stopHandlerBuilder_
);
1985 newRunner
.stopHandlerBuilder_
= std::make_unique
<StopHandlerBuilder
>();
1991 void Mdrunner::BuilderImplementation::addNonBonded(const char* nbpu_opt
)
1993 nbpu_opt_
= nbpu_opt
;
1996 void Mdrunner::BuilderImplementation::addPME(const char* pme_opt
, const char* pme_fft_opt
)
1999 pme_fft_opt_
= pme_fft_opt
;
2002 void Mdrunner::BuilderImplementation::addBondedTaskAssignment(const char* bonded_opt
)
2004 bonded_opt_
= bonded_opt
;
2007 void Mdrunner::BuilderImplementation::addUpdateTaskAssignment(const char* update_opt
)
2009 update_opt_
= update_opt
;
2012 void Mdrunner::BuilderImplementation::addHardwareOptions(const gmx_hw_opt_t
& hardwareOptions
)
2014 hardwareOptions_
= hardwareOptions
;
2017 void Mdrunner::BuilderImplementation::addFilenames(ArrayRef
<const t_filenm
> filenames
)
2019 filenames_
= filenames
;
2022 void Mdrunner::BuilderImplementation::addOutputEnvironment(gmx_output_env_t
* outputEnvironment
)
2024 outputEnvironment_
= outputEnvironment
;
2027 void Mdrunner::BuilderImplementation::addLogFile(t_fileio
* logFileHandle
)
2029 logFileHandle_
= logFileHandle
;
2032 void Mdrunner::BuilderImplementation::addStopHandlerBuilder(std::unique_ptr
<StopHandlerBuilder
> builder
)
2034 stopHandlerBuilder_
= std::move(builder
);
2037 MdrunnerBuilder::MdrunnerBuilder(std::unique_ptr
<MDModules
> mdModules
,
2038 compat::not_null
<SimulationContext
*> context
) :
2039 impl_
{ std::make_unique
<Mdrunner::BuilderImplementation
>(std::move(mdModules
), context
) }
2043 MdrunnerBuilder::~MdrunnerBuilder() = default;
2045 MdrunnerBuilder
& MdrunnerBuilder::addSimulationMethod(const MdrunOptions
& options
,
2046 real forceWarningThreshold
,
2047 const StartingBehavior startingBehavior
)
2049 impl_
->setExtraMdrunOptions(options
, forceWarningThreshold
, startingBehavior
);
2053 MdrunnerBuilder
& MdrunnerBuilder::addDomainDecomposition(const DomdecOptions
& options
)
2055 impl_
->addDomdec(options
);
2059 MdrunnerBuilder
& MdrunnerBuilder::addNeighborList(int nstlist
)
2061 impl_
->addVerletList(nstlist
);
2065 MdrunnerBuilder
& MdrunnerBuilder::addReplicaExchange(const ReplicaExchangeParameters
& params
)
2067 impl_
->addReplicaExchange(params
);
2071 MdrunnerBuilder
& MdrunnerBuilder::addNonBonded(const char* nbpu_opt
)
2073 impl_
->addNonBonded(nbpu_opt
);
2077 MdrunnerBuilder
& MdrunnerBuilder::addElectrostatics(const char* pme_opt
, const char* pme_fft_opt
)
2079 // The builder method may become more general in the future, but in this version,
2080 // parameters for PME electrostatics are both required and the only parameters
2082 if (pme_opt
&& pme_fft_opt
)
2084 impl_
->addPME(pme_opt
, pme_fft_opt
);
2089 gmx::InvalidInputError("addElectrostatics() arguments must be non-null pointers."));
2094 MdrunnerBuilder
& MdrunnerBuilder::addBondedTaskAssignment(const char* bonded_opt
)
2096 impl_
->addBondedTaskAssignment(bonded_opt
);
2100 MdrunnerBuilder
& MdrunnerBuilder::addUpdateTaskAssignment(const char* update_opt
)
2102 impl_
->addUpdateTaskAssignment(update_opt
);
2106 Mdrunner
MdrunnerBuilder::build()
2108 return impl_
->build();
2111 MdrunnerBuilder
& MdrunnerBuilder::addHardwareOptions(const gmx_hw_opt_t
& hardwareOptions
)
2113 impl_
->addHardwareOptions(hardwareOptions
);
2117 MdrunnerBuilder
& MdrunnerBuilder::addFilenames(ArrayRef
<const t_filenm
> filenames
)
2119 impl_
->addFilenames(filenames
);
2123 MdrunnerBuilder
& MdrunnerBuilder::addOutputEnvironment(gmx_output_env_t
* outputEnvironment
)
2125 impl_
->addOutputEnvironment(outputEnvironment
);
2129 MdrunnerBuilder
& MdrunnerBuilder::addLogFile(t_fileio
* logFileHandle
)
2131 impl_
->addLogFile(logFileHandle
);
2135 MdrunnerBuilder
& MdrunnerBuilder::addStopHandlerBuilder(std::unique_ptr
<StopHandlerBuilder
> builder
)
2137 impl_
->addStopHandlerBuilder(std::move(builder
));
2141 MdrunnerBuilder::MdrunnerBuilder(MdrunnerBuilder
&&) noexcept
= default;
2143 MdrunnerBuilder
& MdrunnerBuilder::operator=(MdrunnerBuilder
&&) noexcept
= default;