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, 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_mdlib
57 #include "gromacs/commandline/filenm.h"
58 #include "gromacs/domdec/domdec.h"
59 #include "gromacs/domdec/domdec_struct.h"
60 #include "gromacs/ewald/ewald-utils.h"
61 #include "gromacs/ewald/pme.h"
62 #include "gromacs/fileio/checkpoint.h"
63 #include "gromacs/fileio/oenv.h"
64 #include "gromacs/fileio/tpxio.h"
65 #include "gromacs/gmxlib/network.h"
66 #include "gromacs/gpu_utils/gpu_utils.h"
67 #include "gromacs/hardware/cpuinfo.h"
68 #include "gromacs/hardware/detecthardware.h"
69 #include "gromacs/hardware/printhardware.h"
70 #include "gromacs/listed-forces/disre.h"
71 #include "gromacs/listed-forces/orires.h"
72 #include "gromacs/math/functions.h"
73 #include "gromacs/math/utilities.h"
74 #include "gromacs/math/vec.h"
75 #include "gromacs/mdlib/calc_verletbuf.h"
76 #include "gromacs/mdlib/constr.h"
77 #include "gromacs/mdlib/force.h"
78 #include "gromacs/mdlib/forcerec.h"
79 #include "gromacs/mdlib/gmx_omp_nthreads.h"
80 #include "gromacs/mdlib/integrator.h"
81 #include "gromacs/mdlib/main.h"
82 #include "gromacs/mdlib/md_support.h"
83 #include "gromacs/mdlib/mdatoms.h"
84 #include "gromacs/mdlib/mdrun.h"
85 #include "gromacs/mdlib/minimize.h"
86 #include "gromacs/mdlib/nb_verlet.h"
87 #include "gromacs/mdlib/nbnxn_gpu_data_mgmt.h"
88 #include "gromacs/mdlib/nbnxn_search.h"
89 #include "gromacs/mdlib/nbnxn_tuning.h"
90 #include "gromacs/mdlib/qmmm.h"
91 #include "gromacs/mdlib/sighandler.h"
92 #include "gromacs/mdlib/sim_util.h"
93 #include "gromacs/mdlib/tpi.h"
94 #include "gromacs/mdrunutility/mdmodules.h"
95 #include "gromacs/mdrunutility/threadaffinity.h"
96 #include "gromacs/mdtypes/commrec.h"
97 #include "gromacs/mdtypes/inputrec.h"
98 #include "gromacs/mdtypes/md_enums.h"
99 #include "gromacs/mdtypes/observableshistory.h"
100 #include "gromacs/mdtypes/state.h"
101 #include "gromacs/pbcutil/pbc.h"
102 #include "gromacs/pulling/pull.h"
103 #include "gromacs/pulling/pull_rotation.h"
104 #include "gromacs/taskassignment/decidegpuusage.h"
105 #include "gromacs/taskassignment/resourcedivision.h"
106 #include "gromacs/taskassignment/taskassignment.h"
107 #include "gromacs/taskassignment/usergpuids.h"
108 #include "gromacs/timing/wallcycle.h"
109 #include "gromacs/topology/mtop_util.h"
110 #include "gromacs/trajectory/trajectoryframe.h"
111 #include "gromacs/utility/cstringutil.h"
112 #include "gromacs/utility/exceptions.h"
113 #include "gromacs/utility/fatalerror.h"
114 #include "gromacs/utility/filestream.h"
115 #include "gromacs/utility/gmxassert.h"
116 #include "gromacs/utility/gmxmpi.h"
117 #include "gromacs/utility/logger.h"
118 #include "gromacs/utility/loggerbuilder.h"
119 #include "gromacs/utility/pleasecite.h"
120 #include "gromacs/utility/programcontext.h"
121 #include "gromacs/utility/smalloc.h"
122 #include "gromacs/utility/stringutil.h"
130 #include "corewrap.h"
133 //! First step used in pressure scaling
134 gmx_int64_t deform_init_init_step_tpx
;
135 //! Initial box for pressure scaling
136 matrix deform_init_box_tpx
;
137 //! MPI variable for use in pressure scaling
138 tMPI_Thread_mutex_t deform_init_box_mutex
= TMPI_THREAD_MUTEX_INITIALIZER
;
143 void Mdrunner::reinitializeOnSpawnedThread()
145 // TODO This duplication is formally necessary if any thread might
146 // modify any memory in fnm or the pointers it contains. If the
147 // contents are ever provably const, then we can remove this
148 // allocation (and memory leak).
149 // TODO This should probably become part of a copy constructor for
151 fnm
= dup_tfn(nfile
, fnm
);
153 cr
= reinitialize_commrec_for_this_thread(cr
);
157 // Only the master rank writes to the log files
162 /*! \brief The callback used for running on spawned threads.
164 * Obtains the pointer to the master mdrunner object from the one
165 * argument permitted to the thread-launch API call, copies it to make
166 * a new runner for this thread, reinitializes necessary data, and
167 * proceeds to the simulation. */
168 static void mdrunner_start_fn(void *arg
)
172 auto masterMdrunner
= reinterpret_cast<const gmx::Mdrunner
*>(arg
);
173 /* copy the arg list to make sure that it's thread-local. This
174 doesn't copy pointed-to items, of course, but those are all
176 gmx::Mdrunner mdrunner
= *masterMdrunner
;
177 mdrunner
.reinitializeOnSpawnedThread();
180 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
;
184 /*! \brief Start thread-MPI threads.
186 * Called by mdrunner() to start a specific number of threads
187 * (including the main thread) for thread-parallel runs. This in turn
188 * calls mdrunner() for each thread. All options are the same as for
190 t_commrec
*Mdrunner::spawnThreads(int numThreadsToLaunch
)
193 /* first check whether we even need to start tMPI */
194 if (numThreadsToLaunch
< 2)
199 gmx::Mdrunner spawnedMdrunner
= *this;
200 // TODO This duplication is formally necessary if any thread might
201 // modify any memory in fnm or the pointers it contains. If the
202 // contents are ever provably const, then we can remove this
203 // allocation (and memory leak).
204 // TODO This should probably become part of a copy constructor for
206 spawnedMdrunner
.fnm
= dup_tfn(this->nfile
, fnm
);
209 /* now spawn new threads that start mdrunner_start_fn(), while
210 the main thread returns, we set thread affinity later */
211 if (tMPI_Init_fn(TRUE
, numThreadsToLaunch
, TMPI_AFFINITY_NONE
,
212 mdrunner_start_fn
, static_cast<void*>(&spawnedMdrunner
)) != TMPI_SUCCESS
)
214 GMX_THROW(gmx::InternalError("Failed to spawn thread-MPI threads"));
217 GMX_UNUSED_VALUE(mdrunner_start_fn
);
220 return reinitialize_commrec_for_this_thread(cr
);
225 /*! \brief Initialize variables for Verlet scheme simulation */
226 static void prepare_verlet_scheme(FILE *fplog
,
230 const gmx_mtop_t
*mtop
,
232 bool makeGpuPairList
,
233 const gmx::CpuInfo
&cpuinfo
)
235 /* For NVE simulations, we will retain the initial list buffer */
236 if (EI_DYNAMICS(ir
->eI
) &&
237 ir
->verletbuf_tol
> 0 &&
238 !(EI_MD(ir
->eI
) && ir
->etc
== etcNO
))
240 /* Update the Verlet buffer size for the current run setup */
242 /* Here we assume SIMD-enabled kernels are being used. But as currently
243 * calc_verlet_buffer_size gives the same results for 4x8 and 4x4
244 * and 4x2 gives a larger buffer than 4x4, this is ok.
246 ListSetupType listType
= (makeGpuPairList
? ListSetupType::Gpu
: ListSetupType::CpuSimdWhenSupported
);
247 VerletbufListSetup listSetup
= verletbufGetSafeListSetup(listType
);
250 calc_verlet_buffer_size(mtop
, det(box
), ir
, ir
->nstlist
, ir
->nstlist
- 1, -1, &listSetup
, nullptr, &rlist_new
);
252 if (rlist_new
!= ir
->rlist
)
254 if (fplog
!= nullptr)
256 fprintf(fplog
, "\nChanging rlist from %g to %g for non-bonded %dx%d atom kernels\n\n",
257 ir
->rlist
, rlist_new
,
258 listSetup
.cluster_size_i
, listSetup
.cluster_size_j
);
260 ir
->rlist
= rlist_new
;
264 if (nstlist_cmdline
> 0 && (!EI_DYNAMICS(ir
->eI
) || ir
->verletbuf_tol
<= 0))
266 gmx_fatal(FARGS
, "Can not set nstlist without %s",
267 !EI_DYNAMICS(ir
->eI
) ? "dynamics" : "verlet-buffer-tolerance");
270 if (EI_DYNAMICS(ir
->eI
))
272 /* Set or try nstlist values */
273 increaseNstlist(fplog
, cr
, ir
, nstlist_cmdline
, mtop
, box
, makeGpuPairList
, cpuinfo
);
277 /*! \brief Override the nslist value in inputrec
279 * with value passed on the command line (if any)
281 static void override_nsteps_cmdline(const gmx::MDLogger
&mdlog
,
282 gmx_int64_t nsteps_cmdline
,
287 /* override with anything else than the default -2 */
288 if (nsteps_cmdline
> -2)
290 char sbuf_steps
[STEPSTRSIZE
];
291 char sbuf_msg
[STRLEN
];
293 ir
->nsteps
= nsteps_cmdline
;
294 if (EI_DYNAMICS(ir
->eI
) && nsteps_cmdline
!= -1)
296 sprintf(sbuf_msg
, "Overriding nsteps with value passed on the command line: %s steps, %.3g ps",
297 gmx_step_str(nsteps_cmdline
, sbuf_steps
),
298 fabs(nsteps_cmdline
*ir
->delta_t
));
302 sprintf(sbuf_msg
, "Overriding nsteps with value passed on the command line: %s steps",
303 gmx_step_str(nsteps_cmdline
, sbuf_steps
));
306 GMX_LOG(mdlog
.warning
).asParagraph().appendText(sbuf_msg
);
308 else if (nsteps_cmdline
< -2)
310 gmx_fatal(FARGS
, "Invalid nsteps value passed on the command line: %d",
313 /* Do nothing if nsteps_cmdline == -2 */
319 /*! \brief Return whether GPU acceleration of nonbondeds is useful with the given settings.
321 * If not, logs a message about falling back to CPU code. */
322 static bool gpuAccelerationOfNonbondedIsUseful(const MDLogger
&mdlog
,
323 const t_inputrec
*ir
,
326 if (doRerun
&& ir
->opts
.ngener
> 1)
328 /* Rerun execution time is dominated by I/O and pair search,
329 * so GPUs are not very useful, plus they do not support more
330 * than one energy group. If the user requested GPUs
331 * explicitly, a fatal error is given later. With non-reruns,
332 * we fall back to a single whole-of system energy group
333 * (which runs much faster than a multiple-energy-groups
334 * implementation would), and issue a note in the .log
335 * file. Users can re-run if they want the information. */
336 GMX_LOG(mdlog
.warning
).asParagraph().appendText("Multiple energy groups is not implemented for GPUs, so is not useful for this rerun, so falling back to the CPU");
343 //! \brief Return the correct integrator function.
344 static integrator_t
*my_integrator(unsigned int ei
)
353 if (!EI_DYNAMICS(ei
))
355 GMX_THROW(APIError("do_md integrator would be called for a non-dynamical integrator"));
370 GMX_THROW(APIError("do_tpi integrator would be called for a non-TPI integrator"));
374 GMX_THROW(NotImplementedError("SD2 integrator has been removed"));
376 GMX_THROW(APIError("Non existing integrator selected"));
380 //! Initializes the logger for mdrun.
381 static gmx::LoggerOwner
buildLogger(FILE *fplog
, const t_commrec
*cr
)
383 gmx::LoggerBuilder builder
;
384 if (fplog
!= nullptr)
386 builder
.addTargetFile(gmx::MDLogger::LogLevel::Info
, fplog
);
388 if (cr
== nullptr || SIMMASTER(cr
))
390 builder
.addTargetStream(gmx::MDLogger::LogLevel::Warning
,
391 &gmx::TextOutputFile::standardError());
393 return builder
.build();
396 //! Make a TaskTarget from an mdrun argument string.
397 static TaskTarget
findTaskTarget(const char *optionString
)
399 TaskTarget returnValue
= TaskTarget::Auto
;
401 if (strncmp(optionString
, "auto", 3) == 0)
403 returnValue
= TaskTarget::Auto
;
405 else if (strncmp(optionString
, "cpu", 3) == 0)
407 returnValue
= TaskTarget::Cpu
;
409 else if (strncmp(optionString
, "gpu", 3) == 0)
411 returnValue
= TaskTarget::Gpu
;
415 GMX_ASSERT(false, "Option string should have been checked for sanity already");
421 int Mdrunner::mdrunner()
424 gmx_ddbox_t ddbox
= {0};
425 int npme_major
, npme_minor
;
427 gmx_mtop_t
*mtop
= nullptr;
428 t_forcerec
*fr
= nullptr;
429 t_fcdata
*fcd
= nullptr;
430 real ewaldcoeff_q
= 0;
431 real ewaldcoeff_lj
= 0;
432 gmx_vsite_t
*vsite
= nullptr;
434 int nChargePerturbed
= -1, nTypePerturbed
= 0;
435 gmx_wallcycle_t wcycle
;
436 gmx_walltime_accounting_t walltime_accounting
= nullptr;
438 gmx_int64_t reset_counters
;
439 int nthreads_pme
= 1;
440 gmx_membed_t
* membed
= nullptr;
441 gmx_hw_info_t
*hwinfo
= nullptr;
443 /* CAUTION: threads may be started later on in this function, so
444 cr doesn't reflect the final parallel state right now */
445 gmx::MDModules mdModules
;
446 t_inputrec inputrecInstance
;
447 t_inputrec
*inputrec
= &inputrecInstance
;
450 if (mdrunOptions
.continuationOptions
.appendFiles
)
455 bool doMembed
= opt2bSet("-membed", nfile
, fnm
);
456 bool doRerun
= mdrunOptions
.rerun
;
458 // Handle task-assignment related user options.
459 EmulateGpuNonbonded emulateGpuNonbonded
= (getenv("GMX_EMULATE_GPU") != nullptr ?
460 EmulateGpuNonbonded::Yes
: EmulateGpuNonbonded::No
);
461 std::vector
<int> gpuIdsAvailable
;
464 gpuIdsAvailable
= parseUserGpuIds(hw_opt
.gpuIdsAvailable
);
465 // TODO We could put the GPU IDs into a std::map to find
466 // duplicates, but for the small numbers of IDs involved, this
467 // code is simple and fast.
468 for (size_t i
= 0; i
!= gpuIdsAvailable
.size(); ++i
)
470 for (size_t j
= i
+1; j
!= gpuIdsAvailable
.size(); ++j
)
472 if (gpuIdsAvailable
[i
] == gpuIdsAvailable
[j
])
474 GMX_THROW(InvalidInputError(formatString("The string of available GPU device IDs '%s' may not contain duplicate device IDs", hw_opt
.gpuIdsAvailable
.c_str())));
479 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
;
481 std::vector
<int> userGpuTaskAssignment
;
484 userGpuTaskAssignment
= parseUserGpuIds(hw_opt
.userGpuTaskAssignment
);
486 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
;
487 auto nonbondedTarget
= findTaskTarget(nbpu_opt
);
488 auto pmeTarget
= findTaskTarget(pme_opt
);
489 // TODO find a sensible home and behaviour for this
490 GMX_UNUSED_VALUE(pme_fft_opt
);
491 //auto pmeFftTarget = findTaskTarget(pme_fft_opt);
492 PmeRunMode pmeRunMode
= PmeRunMode::None
;
494 // Here we assume that SIMMASTER(cr) does not change even after the
495 // threads are started.
496 gmx::LoggerOwner
logOwner(buildLogger(fplog
, cr
));
497 gmx::MDLogger
mdlog(logOwner
.logger());
499 hwinfo
= gmx_detect_hardware(mdlog
, cr
);
501 gmx_print_detected_hardware(fplog
, cr
, mdlog
, hwinfo
);
503 std::vector
<int> gpuIdsToUse
;
504 auto compatibleGpus
= getCompatibleGpus(hwinfo
->gpu_info
);
505 if (gpuIdsAvailable
.empty())
507 gpuIdsToUse
= compatibleGpus
;
511 for (const auto &availableGpuId
: gpuIdsAvailable
)
513 bool availableGpuIsCompatible
= false;
514 for (const auto &compatibleGpuId
: compatibleGpus
)
516 if (availableGpuId
== compatibleGpuId
)
518 availableGpuIsCompatible
= true;
522 if (!availableGpuIsCompatible
)
524 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
);
526 gpuIdsToUse
.push_back(availableGpuId
);
530 if (fplog
!= nullptr)
532 /* Print references after all software/hardware printing */
533 please_cite(fplog
, "Abraham2015");
534 please_cite(fplog
, "Pall2015");
535 please_cite(fplog
, "Pronk2013");
536 please_cite(fplog
, "Hess2008b");
537 please_cite(fplog
, "Spoel2005a");
538 please_cite(fplog
, "Lindahl2001a");
539 please_cite(fplog
, "Berendsen95a");
542 std::unique_ptr
<t_state
> globalState
;
546 /* Only the master rank has the global state */
547 globalState
= std::unique_ptr
<t_state
>(new t_state
);
549 /* Read (nearly) all data required for the simulation */
550 read_tpx_state(ftp2fn(efTPR
, nfile
, fnm
), inputrec
, globalState
.get(), mtop
);
552 if (inputrec
->cutoff_scheme
!= ecutsVERLET
)
554 if (nstlist_cmdline
> 0)
556 gmx_fatal(FARGS
, "Can not set nstlist with the group cut-off scheme");
559 if (!compatibleGpus
.empty())
561 GMX_LOG(mdlog
.warning
).asParagraph().appendText(
562 "NOTE: GPU(s) found, but the current simulation can not use GPUs\n"
563 " To use a GPU, set the mdp option: cutoff-scheme = Verlet");
568 /* Check and update the hardware options for internal consistency */
569 check_and_update_hw_opt_1(&hw_opt
, cr
, domdecOptions
.numPmeRanks
);
571 /* Early check for externally set process affinity. */
572 gmx_check_thread_affinity_set(mdlog
, cr
,
573 &hw_opt
, hwinfo
->nthreads_hw_avail
, FALSE
);
575 if (GMX_THREAD_MPI
&& SIMMASTER(cr
))
577 if (domdecOptions
.numPmeRanks
> 0 && hw_opt
.nthreads_tmpi
<= 0)
579 gmx_fatal(FARGS
, "You need to explicitly specify the number of MPI threads (-ntmpi) when using separate PME ranks");
582 /* Since the master knows the cut-off scheme, update hw_opt for this.
583 * This is done later for normal MPI and also once more with tMPI
584 * for all tMPI ranks.
586 check_and_update_hw_opt_2(&hw_opt
, inputrec
->cutoff_scheme
);
588 bool useGpuForNonbonded
= false;
589 bool useGpuForPme
= false;
592 // If the user specified the number of ranks, then we must
593 // respect that, but in default mode, we need to allow for
594 // the number of GPUs to choose the number of ranks.
596 useGpuForNonbonded
= decideWhetherToUseGpusForNonbondedWithThreadMpi
597 (nonbondedTarget
, gpuIdsToUse
, userGpuTaskAssignment
, emulateGpuNonbonded
,
598 inputrec
->cutoff_scheme
== ecutsVERLET
,
599 gpuAccelerationOfNonbondedIsUseful(mdlog
, inputrec
, doRerun
),
600 hw_opt
.nthreads_tmpi
);
601 auto inputSystemHasPme
= EEL_PME(inputrec
->coulombtype
) || EVDW_PME(inputrec
->vdwtype
);
602 auto canUseGpuForPme
= inputSystemHasPme
&& pme_gpu_supports_input(inputrec
, nullptr);
603 useGpuForPme
= decideWhetherToUseGpusForPmeWithThreadMpi
604 (useGpuForNonbonded
, pmeTarget
, gpuIdsToUse
, userGpuTaskAssignment
,
605 canUseGpuForPme
, hw_opt
.nthreads_tmpi
);
607 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
;
608 /* Determine how many thread-MPI ranks to start.
610 * TODO Over-writing the user-supplied value here does
611 * prevent any possible subsequent checks from working
613 hw_opt
.nthreads_tmpi
= get_nthreads_mpi(hwinfo
,
622 // Now start the threads for thread MPI.
623 cr
= spawnThreads(hw_opt
.nthreads_tmpi
);
624 /* The main thread continues here with a new cr. We don't deallocate
625 the old cr because other threads may still be reading it. */
626 // TODO Both master and spawned threads call dup_tfn and
627 // reinitialize_commrec_for_this_thread. Find a way to express
630 /* END OF CAUTION: cr is now reliable */
634 /* now broadcast everything to the non-master nodes/threads: */
635 init_parallel(cr
, inputrec
, mtop
);
638 // Now each rank knows the inputrec that SIMMASTER read and used,
639 // and (if applicable) cr->nnodes has been assigned the number of
640 // thread-MPI ranks that have been chosen. The ranks can now all
641 // run the task-deciding functions and will agree on the result
642 // without needing to communicate.
644 // TODO Should we do the communication in debug mode to support
645 // having an assertion?
647 // Note that these variables describe only their own node.
648 bool useGpuForNonbonded
= false;
649 bool useGpuForPme
= false;
652 useGpuForNonbonded
= decideWhetherToUseGpusForNonbonded(nonbondedTarget
, gpuIdsToUse
, userGpuTaskAssignment
,
653 emulateGpuNonbonded
, inputrec
->cutoff_scheme
== ecutsVERLET
,
654 gpuAccelerationOfNonbondedIsUseful(mdlog
, inputrec
, doRerun
));
655 auto inputSystemHasPme
= EEL_PME(inputrec
->coulombtype
) || EVDW_PME(inputrec
->vdwtype
);
656 auto canUseGpuForPme
= inputSystemHasPme
&& pme_gpu_supports_input(inputrec
, nullptr);
657 useGpuForPme
= decideWhetherToUseGpusForPme(useGpuForNonbonded
, pmeTarget
, userGpuTaskAssignment
, canUseGpuForPme
, cr
->nnodes
);
658 // FIXME decide how to implement -pmefft support
659 pmeRunMode
= (useGpuForPme
? PmeRunMode::GPU
: PmeRunMode::CPU
);
661 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
;
663 // TODO: Error handling
664 mdModules
.assignOptionsToModules(*inputrec
->params
, nullptr);
666 if (fplog
!= nullptr)
668 pr_inputrec(fplog
, 0, "Input Parameters", inputrec
, FALSE
);
669 fprintf(fplog
, "\n");
674 /* now make sure the state is initialized and propagated */
675 set_state_entries(globalState
.get(), inputrec
);
678 /* NM and TPI parallelize over force/energy calculations, not atoms,
679 * so we need to initialize and broadcast the global state.
681 if (inputrec
->eI
== eiNM
|| inputrec
->eI
== eiTPI
)
685 globalState
= std::unique_ptr
<t_state
>(new t_state
);
687 broadcastStateWithoutDynamics(cr
, globalState
.get());
690 /* A parallel command line option consistency check that we can
691 only do after any threads have started. */
692 if (!PAR(cr
) && (domdecOptions
.numCells
[XX
] > 1 ||
693 domdecOptions
.numCells
[YY
] > 1 ||
694 domdecOptions
.numCells
[ZZ
] > 1 ||
695 domdecOptions
.numPmeRanks
> 0))
698 "The -dd or -npme option request a parallel simulation, "
700 "but %s was compiled without threads or MPI enabled"
703 "but the number of MPI-threads (option -ntmpi) is not set or is 1"
705 "but %s was not started through mpirun/mpiexec or only one rank was requested through mpirun/mpiexec"
708 , output_env_get_program_display_name(oenv
)
713 (EI_ENERGY_MINIMIZATION(inputrec
->eI
) || eiNM
== inputrec
->eI
))
715 gmx_fatal(FARGS
, "The .mdp file specified an energy mininization or normal mode algorithm, and these are not compatible with mdrun -rerun");
718 if (can_use_allvsall(inputrec
, TRUE
, cr
, fplog
) && DOMAINDECOMP(cr
))
720 gmx_fatal(FARGS
, "All-vs-all loops do not work with domain decomposition, use a single MPI rank");
723 if (!(EEL_PME(inputrec
->coulombtype
) || EVDW_PME(inputrec
->vdwtype
)))
725 if (domdecOptions
.numPmeRanks
> 0)
727 gmx_fatal_collective(FARGS
, cr
->mpi_comm_mysim
, MASTER(cr
),
728 "PME-only ranks are requested, but the system does not use PME for electrostatics or LJ");
731 domdecOptions
.numPmeRanks
= 0;
734 if (useGpuForNonbonded
&& domdecOptions
.numPmeRanks
< 0)
736 /* With GPUs we don't automatically use PME-only ranks. PME ranks can
737 * improve performance with many threads per GPU, since our OpenMP
738 * scaling is bad, but it's difficult to automate the setup.
740 domdecOptions
.numPmeRanks
= 0;
744 if (domdecOptions
.numPmeRanks
< 0)
746 domdecOptions
.numPmeRanks
= 0; // separate GPU ranks not supported
750 GMX_RELEASE_ASSERT(domdecOptions
.numPmeRanks
== 0, "Separate PME GPU ranks are not yet supported");
757 fcRegisterSteps(inputrec
->nsteps
, inputrec
->init_step
);
761 /* NMR restraints must be initialized before load_checkpoint,
762 * since with time averaging the history is added to t_state.
763 * For proper consistency check we therefore need to extend
765 * So the PME-only nodes (if present) will also initialize
766 * the distance restraints.
770 /* This needs to be called before read_checkpoint to extend the state */
771 init_disres(fplog
, mtop
, inputrec
, cr
, fcd
, globalState
.get(), replExParams
.exchangeInterval
> 0);
773 init_orires(fplog
, mtop
, inputrec
, cr
, globalState
.get(), &(fcd
->orires
));
775 if (inputrecDeform(inputrec
))
777 /* Store the deform reference box before reading the checkpoint */
780 copy_mat(globalState
->box
, box
);
784 gmx_bcast(sizeof(box
), box
, cr
);
786 /* Because we do not have the update struct available yet
787 * in which the reference values should be stored,
788 * we store them temporarily in static variables.
789 * This should be thread safe, since they are only written once
790 * and with identical values.
792 tMPI_Thread_mutex_lock(&deform_init_box_mutex
);
793 deform_init_init_step_tpx
= inputrec
->init_step
;
794 copy_mat(box
, deform_init_box_tpx
);
795 tMPI_Thread_mutex_unlock(&deform_init_box_mutex
);
798 ObservablesHistory observablesHistory
= {};
800 ContinuationOptions
&continuationOptions
= mdrunOptions
.continuationOptions
;
802 if (continuationOptions
.startedFromCheckpoint
)
804 /* Check if checkpoint file exists before doing continuation.
805 * This way we can use identical input options for the first and subsequent runs...
809 load_checkpoint(opt2fn_master("-cpi", nfile
, fnm
, cr
), &fplog
,
810 cr
, domdecOptions
.numCells
,
811 inputrec
, globalState
.get(),
812 &bReadEkin
, &observablesHistory
,
813 continuationOptions
.appendFiles
,
814 continuationOptions
.appendFilesOptionSet
,
815 mdrunOptions
.reproducible
);
819 continuationOptions
.haveReadEkin
= true;
823 if (SIMMASTER(cr
) && continuationOptions
.appendFiles
)
825 gmx_log_open(ftp2fn(efLOG
, nfile
, fnm
), cr
,
826 continuationOptions
.appendFiles
, &fplog
);
827 logOwner
= buildLogger(fplog
, nullptr);
828 mdlog
= logOwner
.logger();
831 /* override nsteps with value set on the commamdline */
832 override_nsteps_cmdline(mdlog
, mdrunOptions
.numStepsCommandline
, inputrec
);
836 copy_mat(globalState
->box
, box
);
841 gmx_bcast(sizeof(box
), box
, cr
);
844 /* Update rlist and nstlist. */
845 if (inputrec
->cutoff_scheme
== ecutsVERLET
)
847 prepare_verlet_scheme(fplog
, cr
, inputrec
, nstlist_cmdline
, mtop
, box
,
848 useGpuForNonbonded
|| (emulateGpuNonbonded
== EmulateGpuNonbonded::Yes
), *hwinfo
->cpuInfo
);
851 if (PAR(cr
) && !(EI_TPI(inputrec
->eI
) ||
852 inputrec
->eI
== eiNM
))
854 const rvec
*xOnMaster
= (SIMMASTER(cr
) ? as_rvec_array(globalState
->x
.data()) : nullptr);
856 cr
->dd
= init_domain_decomposition(fplog
, cr
, domdecOptions
, mdrunOptions
,
859 &ddbox
, &npme_major
, &npme_minor
);
860 // Note that local state still does not exist yet.
864 /* PME, if used, is done on all nodes with 1D decomposition */
866 cr
->duty
= (DUTY_PP
| DUTY_PME
);
870 if (inputrec
->ePBC
== epbcSCREW
)
873 "pbc=%s is only implemented with domain decomposition",
874 epbc_names
[inputrec
->ePBC
]);
880 /* After possible communicator splitting in make_dd_communicators.
881 * we can set up the intra/inter node communication.
883 gmx_setup_nodecomm(fplog
, cr
);
886 /* Initialize per-physical-node MPI process/thread ID and counters. */
887 gmx_init_intranode_counters(cr
);
891 GMX_LOG(mdlog
.warning
).asParagraph().appendTextFormatted(
892 "This is simulation %d out of %d running as a composite GROMACS\n"
893 "multi-simulation job. Setup for this simulation:\n",
894 cr
->ms
->sim
, cr
->ms
->nsim
);
896 GMX_LOG(mdlog
.warning
).appendTextFormatted(
900 cr
->nnodes
== 1 ? "thread" : "threads"
902 cr
->nnodes
== 1 ? "process" : "processes"
908 /* Check and update hw_opt for the cut-off scheme */
909 check_and_update_hw_opt_2(&hw_opt
, inputrec
->cutoff_scheme
);
911 /* Check and update the number of OpenMP threads requested */
912 checkAndUpdateRequestedNumOpenmpThreads(&hw_opt
, *hwinfo
, cr
, pmeRunMode
, *mtop
);
914 gmx_omp_nthreads_init(mdlog
, cr
,
915 hwinfo
->nthreads_hw_avail
,
917 hw_opt
.nthreads_omp_pme
,
918 !thisRankHasDuty(cr
, DUTY_PP
),
919 inputrec
->cutoff_scheme
== ecutsVERLET
);
922 if (EI_TPI(inputrec
->eI
) &&
923 inputrec
->cutoff_scheme
== ecutsVERLET
)
925 gmx_feenableexcept();
929 // Build a data structure that expresses which kinds of non-bonded
930 // task are handled by this rank.
932 // TODO Later, this might become a loop over all registered modules
933 // relevant to the mdp inputs, to find those that have such tasks.
935 // TODO This could move before init_domain_decomposition() as part
936 // of refactoring that separates the responsibility for duty
937 // assignment from setup for communication between tasks, and
938 // setup for tasks handled with a domain (ie including short-ranged
939 // tasks, bonded tasks, etc.).
941 // Note that in general useGpuForNonbonded, etc. can have a value
942 // that is inconsistent with the presence of actual GPUs on any
943 // rank, and that is not known to be a problem until the
944 // duty of the ranks on a node become node.
946 // TODO Later we might need the concept of computeTasksOnThisRank,
947 // from which we construct gpuTasksOnThisRank.
949 // Currently the DD code assigns duty to ranks that can
950 // include PP work that currently can be executed on a single
951 // GPU, if present and compatible. This has to be coordinated
952 // across PP ranks on a node, with possible multiple devices
953 // or sharing devices on a node, either from the user
954 // selection, or automatically.
955 auto haveGpus
= !gpuIdsToUse
.empty();
956 std::vector
<GpuTask
> gpuTasksOnThisRank
;
957 if (thisRankHasDuty(cr
, DUTY_PP
))
959 if (useGpuForNonbonded
)
963 gpuTasksOnThisRank
.push_back(GpuTask::Nonbonded
);
965 else if (nonbondedTarget
== TaskTarget::Gpu
)
967 gmx_fatal(FARGS
, "Cannot run short-ranged nonbonded interactions on a GPU because there is none detected.");
971 // TODO cr->duty & DUTY_PME should imply that a PME algorithm is active, but currently does not.
972 if (EEL_PME(inputrec
->coulombtype
) && (thisRankHasDuty(cr
, DUTY_PME
)))
978 gpuTasksOnThisRank
.push_back(GpuTask::Pme
);
980 else if (pmeTarget
== TaskTarget::Gpu
)
982 gmx_fatal(FARGS
, "Cannot run PME on a GPU because there is none detected.");
987 GpuTaskAssignment gpuTaskAssignment
;
990 // Produce the task assignment for this rank.
991 gpuTaskAssignment
= runTaskAssignment(gpuIdsToUse
, userGpuTaskAssignment
, *hwinfo
,
992 mdlog
, cr
, gpuTasksOnThisRank
);
994 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
;
996 /* Prevent other ranks from continuing after an issue was found
997 * and reported as a fatal error.
999 * TODO This function implements a barrier so that MPI runtimes
1000 * can organize an orderly shutdown if one of the ranks has had to
1001 * issue a fatal error in various code already run. When we have
1002 * MPI-aware error handling and reporting, this should be
1007 MPI_Barrier(cr
->mpi_comm_mysim
);
1011 MPI_Barrier(cr
->ms
->mpi_comm_masters
);
1015 /* Now that we know the setup is consistent, check for efficiency */
1016 check_resource_division_efficiency(hwinfo
, hw_opt
.nthreads_tot
, !gpuTaskAssignment
.empty(), mdrunOptions
.ntompOptionIsSet
,
1019 gmx_device_info_t
*nonbondedDeviceInfo
= nullptr;
1021 if (thisRankHasDuty(cr
, DUTY_PP
))
1023 // This works because only one task of each type is currently permitted.
1024 auto nbGpuTaskMapping
= std::find_if(gpuTaskAssignment
.begin(), gpuTaskAssignment
.end(),
1025 hasTaskType
<GpuTask::Nonbonded
>);
1026 if (nbGpuTaskMapping
!= gpuTaskAssignment
.end())
1028 int nonbondedDeviceId
= nbGpuTaskMapping
->deviceId_
;
1029 nonbondedDeviceInfo
= getDeviceInfo(hwinfo
->gpu_info
, nonbondedDeviceId
);
1030 init_gpu(mdlog
, nonbondedDeviceInfo
);
1032 if (DOMAINDECOMP(cr
))
1034 /* When we share GPUs over ranks, we need to know this for the DLB */
1035 dd_setup_dlb_resource_sharing(cr
, nonbondedDeviceId
);
1041 gmx_device_info_t
*pmeDeviceInfo
= nullptr;
1042 // This works because only one task of each type is currently permitted.
1043 auto pmeGpuTaskMapping
= std::find_if(gpuTaskAssignment
.begin(), gpuTaskAssignment
.end(), hasTaskType
<GpuTask::Pme
>);
1044 if (pmeGpuTaskMapping
!= gpuTaskAssignment
.end())
1046 pmeDeviceInfo
= getDeviceInfo(hwinfo
->gpu_info
, pmeGpuTaskMapping
->deviceId_
);
1049 /* getting number of PP/PME threads
1050 PME: env variable should be read only on one node to make sure it is
1051 identical everywhere;
1053 nthreads_pme
= gmx_omp_nthreads_get(emntPME
);
1055 wcycle
= wallcycle_init(fplog
, mdrunOptions
.timingOptions
.resetStep
, cr
);
1059 /* Master synchronizes its value of reset_counters with all nodes
1060 * including PME only nodes */
1061 reset_counters
= wcycle_get_reset_counters(wcycle
);
1062 gmx_bcast_sim(sizeof(reset_counters
), &reset_counters
, cr
);
1063 wcycle_set_reset_counters(wcycle
, reset_counters
);
1066 // Membrane embedding must be initialized before we call init_forcerec()
1071 fprintf(stderr
, "Initializing membed");
1073 /* Note that membed cannot work in parallel because mtop is
1074 * changed here. Fix this if we ever want to make it run with
1075 * multiple ranks. */
1076 membed
= init_membed(fplog
, nfile
, fnm
, mtop
, inputrec
, globalState
.get(), cr
, &mdrunOptions
.checkpointOptions
.period
);
1079 std::unique_ptr
<MDAtoms
> mdAtoms
;
1082 if (thisRankHasDuty(cr
, DUTY_PP
))
1084 /* Initiate forcerecord */
1086 fr
->forceProviders
= mdModules
.initForceProviders();
1087 init_forcerec(fplog
, mdlog
, fr
, fcd
,
1088 inputrec
, mtop
, cr
, box
,
1089 opt2fn("-table", nfile
, fnm
),
1090 opt2fn("-tablep", nfile
, fnm
),
1091 getFilenm("-tableb", nfile
, fnm
),
1092 nonbondedDeviceInfo
,
1096 /* Initialize QM-MM */
1099 init_QMMMrec(cr
, mtop
, inputrec
, fr
);
1102 /* Initialize the mdAtoms structure.
1103 * mdAtoms is not filled with atom data,
1104 * as this can not be done now with domain decomposition.
1106 const bool useGpuForPme
= (pmeRunMode
== PmeRunMode::GPU
) || (pmeRunMode
== PmeRunMode::Hybrid
);
1107 mdAtoms
= makeMDAtoms(fplog
, *mtop
, *inputrec
, useGpuForPme
&& thisRankHasDuty(cr
, DUTY_PME
));
1110 // The pinning of coordinates in the global state object works, because we only use
1111 // PME on GPU without DD or on a separate PME rank, and because the local state pointer
1112 // points to the global state object without DD.
1113 // FIXME: MD and EM separately set up the local state - this should happen in the same function,
1114 // which should also perform the pinning.
1115 changePinningPolicy(&globalState
->x
, useGpuForPme
? PinningPolicy::CanBePinned
: PinningPolicy::CannotBePinned
);
1118 /* Initialize the virtual site communication */
1119 vsite
= initVsite(*mtop
, cr
);
1121 calc_shifts(box
, fr
->shift_vec
);
1123 /* With periodic molecules the charge groups should be whole at start up
1124 * and the virtual sites should not be far from their proper positions.
1126 if (!inputrec
->bContinuation
&& MASTER(cr
) &&
1127 !(inputrec
->ePBC
!= epbcNONE
&& inputrec
->bPeriodicMols
))
1129 /* Make molecules whole at start of run */
1130 if (fr
->ePBC
!= epbcNONE
)
1132 rvec
*xGlobal
= as_rvec_array(globalState
->x
.data());
1133 do_pbc_first_mtop(fplog
, inputrec
->ePBC
, box
, mtop
, xGlobal
);
1137 /* Correct initial vsite positions are required
1138 * for the initial distribution in the domain decomposition
1139 * and for the initial shell prediction.
1141 constructVsitesGlobal(*mtop
, globalState
->x
);
1145 if (EEL_PME(fr
->ic
->eeltype
) || EVDW_PME(fr
->ic
->vdwtype
))
1147 ewaldcoeff_q
= fr
->ic
->ewaldcoeff_q
;
1148 ewaldcoeff_lj
= fr
->ic
->ewaldcoeff_lj
;
1153 /* This is a PME only node */
1155 GMX_ASSERT(globalState
== nullptr, "We don't need the state on a PME only rank and expect it to be unitialized");
1157 ewaldcoeff_q
= calc_ewaldcoeff_q(inputrec
->rcoulomb
, inputrec
->ewald_rtol
);
1158 ewaldcoeff_lj
= calc_ewaldcoeff_lj(inputrec
->rvdw
, inputrec
->ewald_rtol_lj
);
1161 gmx_pme_t
*sepPmeData
= nullptr;
1162 // This reference hides the fact that PME data is owned by runner on PME-only ranks and by forcerec on other ranks
1163 GMX_ASSERT(thisRankHasDuty(cr
, DUTY_PP
) == (fr
!= nullptr), "Double-checking that only PME-only ranks have no forcerec");
1164 gmx_pme_t
* &pmedata
= fr
? fr
->pmedata
: sepPmeData
;
1166 if (hw_opt
.thread_affinity
!= threadaffOFF
)
1168 /* Before setting affinity, check whether the affinity has changed
1169 * - which indicates that probably the OpenMP library has changed it
1170 * since we first checked).
1172 gmx_check_thread_affinity_set(mdlog
, cr
,
1173 &hw_opt
, hwinfo
->nthreads_hw_avail
, TRUE
);
1176 /* threads on this MPI process or TMPI thread */
1177 if (thisRankHasDuty(cr
, DUTY_PP
))
1179 nthread_local
= gmx_omp_nthreads_get(emntNonbonded
);
1183 nthread_local
= gmx_omp_nthreads_get(emntPME
);
1186 /* Set the CPU affinity */
1187 gmx_set_thread_affinity(mdlog
, cr
, &hw_opt
, *hwinfo
->hardwareTopology
,
1188 nthread_local
, nullptr);
1191 /* Initiate PME if necessary,
1192 * either on all nodes or on dedicated PME nodes only. */
1193 if (EEL_PME(inputrec
->coulombtype
) || EVDW_PME(inputrec
->vdwtype
))
1195 if (mdAtoms
&& mdAtoms
->mdatoms())
1197 nChargePerturbed
= mdAtoms
->mdatoms()->nChargePerturbed
;
1198 if (EVDW_PME(inputrec
->vdwtype
))
1200 nTypePerturbed
= mdAtoms
->mdatoms()->nTypePerturbed
;
1203 if (cr
->npmenodes
> 0)
1205 /* The PME only nodes need to know nChargePerturbed(FEP on Q) and nTypePerturbed(FEP on LJ)*/
1206 gmx_bcast_sim(sizeof(nChargePerturbed
), &nChargePerturbed
, cr
);
1207 gmx_bcast_sim(sizeof(nTypePerturbed
), &nTypePerturbed
, cr
);
1210 if (thisRankHasDuty(cr
, DUTY_PME
))
1214 if (pmeDeviceInfo
!= nullptr && pmeDeviceInfo
!= nonbondedDeviceInfo
)
1216 GMX_THROW(NotImplementedError
1217 ("PME on a GPU can run only on the same GPU as nonbonded, because "
1218 "context switching is not yet supported."));
1220 pmedata
= gmx_pme_init(cr
, npme_major
, npme_minor
, inputrec
,
1221 mtop
? mtop
->natoms
: 0, nChargePerturbed
, nTypePerturbed
,
1222 mdrunOptions
.reproducible
,
1223 ewaldcoeff_q
, ewaldcoeff_lj
,
1225 pmeRunMode
, nullptr, pmeDeviceInfo
, mdlog
);
1227 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
;
1232 if (EI_DYNAMICS(inputrec
->eI
))
1234 /* Turn on signal handling on all nodes */
1236 * (A user signal from the PME nodes (if any)
1237 * is communicated to the PP nodes.
1239 signal_handler_install();
1242 if (thisRankHasDuty(cr
, DUTY_PP
))
1244 /* Assumes uniform use of the number of OpenMP threads */
1245 walltime_accounting
= walltime_accounting_init(gmx_omp_nthreads_get(emntDefault
));
1247 if (inputrec
->bPull
)
1249 /* Initialize pull code */
1250 inputrec
->pull_work
=
1251 init_pull(fplog
, inputrec
->pull
, inputrec
, nfile
, fnm
,
1252 mtop
, cr
, oenv
, inputrec
->fepvals
->init_lambda
,
1253 EI_DYNAMICS(inputrec
->eI
) && MASTER(cr
),
1254 continuationOptions
);
1259 /* Initialize enforced rotation code */
1260 init_rot(fplog
, inputrec
, nfile
, fnm
, cr
, globalState
.get(), mtop
, oenv
, mdrunOptions
);
1263 /* Let init_constraints know whether we have essential dynamics constraints.
1264 * TODO: inputrec should tell us whether we use an algorithm, not a file option or the checkpoint
1266 bool doEdsam
= (opt2fn_null("-ei", nfile
, fnm
) != nullptr || observablesHistory
.edsamHistory
);
1268 constr
= init_constraints(fplog
, mtop
, inputrec
, doEdsam
, cr
);
1270 if (DOMAINDECOMP(cr
))
1272 GMX_RELEASE_ASSERT(fr
, "fr was NULL while cr->duty was DUTY_PP");
1273 /* This call is not included in init_domain_decomposition mainly
1274 * because fr->cginfo_mb is set later.
1276 dd_init_bondeds(fplog
, cr
->dd
, mtop
, vsite
, inputrec
,
1277 domdecOptions
.checkBondedInteractions
,
1281 /* Now do whatever the user wants us to do (how flexible...) */
1282 my_integrator(inputrec
->eI
) (fplog
, cr
, mdlog
, nfile
, fnm
,
1286 mdModules
.outputProvider(),
1290 &observablesHistory
,
1291 mdAtoms
.get(), nrnb
, wcycle
, fr
,
1294 walltime_accounting
);
1298 finish_rot(inputrec
->rot
);
1301 if (inputrec
->bPull
)
1303 finish_pull(inputrec
->pull_work
);
1309 GMX_RELEASE_ASSERT(pmedata
, "pmedata was NULL while cr->duty was not DUTY_PP");
1311 walltime_accounting
= walltime_accounting_init(gmx_omp_nthreads_get(emntPME
));
1312 gmx_pmeonly(pmedata
, cr
, nrnb
, wcycle
, walltime_accounting
, inputrec
, pmeRunMode
);
1315 wallcycle_stop(wcycle
, ewcRUN
);
1317 /* Finish up, write some stuff
1318 * if rerunMD, don't write last frame again
1320 finish_run(fplog
, mdlog
, cr
,
1321 inputrec
, nrnb
, wcycle
, walltime_accounting
,
1322 fr
? fr
->nbv
: nullptr,
1324 EI_DYNAMICS(inputrec
->eI
) && !MULTISIM(cr
));
1329 gmx_pme_destroy(pmedata
);
1333 // FIXME: this is only here to manually unpin mdAtoms->chargeA_ and state->x,
1334 // before we destroy the GPU context(s) in free_gpu_resources().
1335 // Pinned buffers are associated with contexts in CUDA.
1336 // As soon as we destroy GPU contexts after mdrunner() exits, these lines should go.
1337 mdAtoms
.reset(nullptr);
1338 globalState
.reset(nullptr);
1340 /* Free GPU memory and set a physical node tMPI barrier (which should eventually go away) */
1341 free_gpu_resources(fr
, cr
);
1342 free_gpu(nonbondedDeviceInfo
);
1346 free_membed(membed
);
1349 gmx_hardware_info_free();
1351 /* Does what it says */
1352 print_date_and_time(fplog
, cr
->nodeid
, "Finished mdrun", gmx_gettime());
1353 walltime_accounting_destroy(walltime_accounting
);
1355 /* Close logfile already here if we were appending to it */
1356 if (MASTER(cr
) && continuationOptions
.appendFiles
)
1358 gmx_log_close(fplog
);
1361 rc
= (int)gmx_get_stop_condition();
1364 /* we need to join all threads. The sub-threads join when they
1365 exit this function, but the master thread needs to be told to
1367 if (PAR(cr
) && MASTER(cr
))