Fix task assignment with multiruns
[alexxy/gromacs.git] / src / gromacs / mdrun / runner.cpp
1 /*
2  * This file is part of the GROMACS molecular simulation package.
3  *
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.
10  *
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.
15  *
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.
20  *
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.
25  *
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.
33  *
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.
36  */
37 /*! \internal \file
38  *
39  * \brief Implements the MD runner routine calling all integrators.
40  *
41  * \author David van der Spoel <david.vanderspoel@icm.uu.se>
42  * \ingroup module_mdrun
43  */
44 #include "gmxpre.h"
45
46 #include "runner.h"
47
48 #include "config.h"
49
50 #include <cassert>
51 #include <cinttypes>
52 #include <csignal>
53 #include <cstdlib>
54 #include <cstring>
55
56 #include <algorithm>
57 #include <memory>
58
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/device_context.h"
77 #include "gromacs/gpu_utils/device_stream_manager.h"
78 #include "gromacs/hardware/cpuinfo.h"
79 #include "gromacs/hardware/detecthardware.h"
80 #include "gromacs/hardware/device_management.h"
81 #include "gromacs/hardware/printhardware.h"
82 #include "gromacs/imd/imd.h"
83 #include "gromacs/listed_forces/disre.h"
84 #include "gromacs/listed_forces/gpubonded.h"
85 #include "gromacs/listed_forces/listed_forces.h"
86 #include "gromacs/listed_forces/orires.h"
87 #include "gromacs/math/functions.h"
88 #include "gromacs/math/utilities.h"
89 #include "gromacs/math/vec.h"
90 #include "gromacs/mdlib/boxdeformation.h"
91 #include "gromacs/mdlib/broadcaststructs.h"
92 #include "gromacs/mdlib/calc_verletbuf.h"
93 #include "gromacs/mdlib/dispersioncorrection.h"
94 #include "gromacs/mdlib/enerdata_utils.h"
95 #include "gromacs/mdlib/force.h"
96 #include "gromacs/mdlib/forcerec.h"
97 #include "gromacs/mdlib/gmx_omp_nthreads.h"
98 #include "gromacs/mdlib/makeconstraints.h"
99 #include "gromacs/mdlib/md_support.h"
100 #include "gromacs/mdlib/mdatoms.h"
101 #include "gromacs/mdlib/sighandler.h"
102 #include "gromacs/mdlib/stophandler.h"
103 #include "gromacs/mdlib/tgroup.h"
104 #include "gromacs/mdlib/updategroups.h"
105 #include "gromacs/mdlib/vsite.h"
106 #include "gromacs/mdrun/mdmodules.h"
107 #include "gromacs/mdrun/simulationcontext.h"
108 #include "gromacs/mdrun/simulationinput.h"
109 #include "gromacs/mdrun/simulationinputhandle.h"
110 #include "gromacs/mdrunutility/handlerestart.h"
111 #include "gromacs/mdrunutility/logging.h"
112 #include "gromacs/mdrunutility/multisim.h"
113 #include "gromacs/mdrunutility/printtime.h"
114 #include "gromacs/mdrunutility/threadaffinity.h"
115 #include "gromacs/mdtypes/checkpointdata.h"
116 #include "gromacs/mdtypes/commrec.h"
117 #include "gromacs/mdtypes/enerdata.h"
118 #include "gromacs/mdtypes/fcdata.h"
119 #include "gromacs/mdtypes/forcerec.h"
120 #include "gromacs/mdtypes/group.h"
121 #include "gromacs/mdtypes/inputrec.h"
122 #include "gromacs/mdtypes/interaction_const.h"
123 #include "gromacs/mdtypes/md_enums.h"
124 #include "gromacs/mdtypes/mdatom.h"
125 #include "gromacs/mdtypes/mdrunoptions.h"
126 #include "gromacs/mdtypes/observableshistory.h"
127 #include "gromacs/mdtypes/simulation_workload.h"
128 #include "gromacs/mdtypes/state.h"
129 #include "gromacs/mdtypes/state_propagator_data_gpu.h"
130 #include "gromacs/modularsimulator/modularsimulator.h"
131 #include "gromacs/nbnxm/gpu_data_mgmt.h"
132 #include "gromacs/nbnxm/nbnxm.h"
133 #include "gromacs/nbnxm/pairlist_tuning.h"
134 #include "gromacs/pbcutil/pbc.h"
135 #include "gromacs/pulling/output.h"
136 #include "gromacs/pulling/pull.h"
137 #include "gromacs/pulling/pull_rotation.h"
138 #include "gromacs/restraint/manager.h"
139 #include "gromacs/restraint/restraintmdmodule.h"
140 #include "gromacs/restraint/restraintpotential.h"
141 #include "gromacs/swap/swapcoords.h"
142 #include "gromacs/taskassignment/decidegpuusage.h"
143 #include "gromacs/taskassignment/decidesimulationworkload.h"
144 #include "gromacs/taskassignment/resourcedivision.h"
145 #include "gromacs/taskassignment/taskassignment.h"
146 #include "gromacs/taskassignment/usergpuids.h"
147 #include "gromacs/timing/gpu_timing.h"
148 #include "gromacs/timing/wallcycle.h"
149 #include "gromacs/timing/wallcyclereporting.h"
150 #include "gromacs/topology/mtop_util.h"
151 #include "gromacs/trajectory/trajectoryframe.h"
152 #include "gromacs/utility/basenetwork.h"
153 #include "gromacs/utility/cstringutil.h"
154 #include "gromacs/utility/exceptions.h"
155 #include "gromacs/utility/fatalerror.h"
156 #include "gromacs/utility/filestream.h"
157 #include "gromacs/utility/gmxassert.h"
158 #include "gromacs/utility/gmxmpi.h"
159 #include "gromacs/utility/keyvaluetree.h"
160 #include "gromacs/utility/logger.h"
161 #include "gromacs/utility/loggerbuilder.h"
162 #include "gromacs/utility/mdmodulenotification.h"
163 #include "gromacs/utility/physicalnodecommunicator.h"
164 #include "gromacs/utility/pleasecite.h"
165 #include "gromacs/utility/programcontext.h"
166 #include "gromacs/utility/smalloc.h"
167 #include "gromacs/utility/stringutil.h"
168
169 #include "isimulator.h"
170 #include "membedholder.h"
171 #include "replicaexchange.h"
172 #include "simulatorbuilder.h"
173
174 #if GMX_FAHCORE
175 #    include "corewrap.h"
176 #endif
177
178 namespace gmx
179 {
180
181
182 /*! \brief Manage any development feature flag variables encountered
183  *
184  * The use of dev features indicated by environment variables is
185  * logged in order to ensure that runs with such features enabled can
186  * be identified from their log and standard output. Any cross
187  * dependencies are also checked, and if unsatisfied, a fatal error
188  * issued.
189  *
190  * Note that some development features overrides are applied already here:
191  * the GPU communication flags are set to false in non-tMPI and non-CUDA builds.
192  *
193  * \param[in]  mdlog                Logger object.
194  * \param[in]  useGpuForNonbonded   True if the nonbonded task is offloaded in this run.
195  * \param[in]  pmeRunMode           The PME run mode for this run
196  * \returns                         The object populated with development feature flags.
197  */
198 static DevelopmentFeatureFlags manageDevelopmentFeatures(const gmx::MDLogger& mdlog,
199                                                          const bool           useGpuForNonbonded,
200                                                          const PmeRunMode     pmeRunMode)
201 {
202     DevelopmentFeatureFlags devFlags;
203
204     // Some builds of GCC 5 give false positive warnings that these
205     // getenv results are ignored when clearly they are used.
206 #pragma GCC diagnostic push
207 #pragma GCC diagnostic ignored "-Wunused-result"
208
209     devFlags.enableGpuBufferOps =
210             GMX_GPU_CUDA && useGpuForNonbonded && (getenv("GMX_USE_GPU_BUFFER_OPS") != nullptr);
211     devFlags.enableGpuHaloExchange = GMX_GPU_CUDA && GMX_THREAD_MPI && getenv("GMX_GPU_DD_COMMS") != nullptr;
212     devFlags.forceGpuUpdateDefault = (getenv("GMX_FORCE_UPDATE_DEFAULT_GPU") != nullptr);
213     devFlags.enableGpuPmePPComm =
214             GMX_GPU_CUDA && GMX_THREAD_MPI && getenv("GMX_GPU_PME_PP_COMMS") != nullptr;
215
216 #pragma GCC diagnostic pop
217
218     if (devFlags.enableGpuBufferOps)
219     {
220         GMX_LOG(mdlog.warning)
221                 .asParagraph()
222                 .appendTextFormatted(
223                         "This run uses the 'GPU buffer ops' feature, enabled by the "
224                         "GMX_USE_GPU_BUFFER_OPS environment variable.");
225     }
226
227     if (devFlags.forceGpuUpdateDefault)
228     {
229         GMX_LOG(mdlog.warning)
230                 .asParagraph()
231                 .appendTextFormatted(
232                         "This run will default to '-update gpu' as requested by the "
233                         "GMX_FORCE_UPDATE_DEFAULT_GPU environment variable. GPU update with domain "
234                         "decomposition lacks substantial testing and should be used with caution.");
235     }
236
237     if (devFlags.enableGpuHaloExchange)
238     {
239         if (useGpuForNonbonded)
240         {
241             if (!devFlags.enableGpuBufferOps)
242             {
243                 GMX_LOG(mdlog.warning)
244                         .asParagraph()
245                         .appendTextFormatted(
246                                 "Enabling GPU buffer operations required by GMX_GPU_DD_COMMS "
247                                 "(equivalent with GMX_USE_GPU_BUFFER_OPS=1).");
248                 devFlags.enableGpuBufferOps = true;
249             }
250             GMX_LOG(mdlog.warning)
251                     .asParagraph()
252                     .appendTextFormatted(
253                             "This run has requested the 'GPU halo exchange' feature, enabled by "
254                             "the "
255                             "GMX_GPU_DD_COMMS environment variable.");
256         }
257         else
258         {
259             GMX_LOG(mdlog.warning)
260                     .asParagraph()
261                     .appendTextFormatted(
262                             "GMX_GPU_DD_COMMS environment variable detected, but the 'GPU "
263                             "halo exchange' feature will not be enabled as nonbonded interactions "
264                             "are not offloaded.");
265             devFlags.enableGpuHaloExchange = false;
266         }
267     }
268
269     if (devFlags.enableGpuPmePPComm)
270     {
271         if (pmeRunMode == PmeRunMode::GPU)
272         {
273             if (!devFlags.enableGpuBufferOps)
274             {
275                 GMX_LOG(mdlog.warning)
276                         .asParagraph()
277                         .appendTextFormatted(
278                                 "Enabling GPU buffer operations required by GMX_GPU_PME_PP_COMMS "
279                                 "(equivalent with GMX_USE_GPU_BUFFER_OPS=1).");
280                 devFlags.enableGpuBufferOps = true;
281             }
282             GMX_LOG(mdlog.warning)
283                     .asParagraph()
284                     .appendTextFormatted(
285                             "This run uses the 'GPU PME-PP communications' feature, enabled "
286                             "by the GMX_GPU_PME_PP_COMMS environment variable.");
287         }
288         else
289         {
290             std::string clarification;
291             if (pmeRunMode == PmeRunMode::Mixed)
292             {
293                 clarification =
294                         "PME FFT and gather are not offloaded to the GPU (PME is running in mixed "
295                         "mode).";
296             }
297             else
298             {
299                 clarification = "PME is not offloaded to the GPU.";
300             }
301             GMX_LOG(mdlog.warning)
302                     .asParagraph()
303                     .appendText(
304                             "GMX_GPU_PME_PP_COMMS environment variable detected, but the "
305                             "'GPU PME-PP communications' feature was not enabled as "
306                             + clarification);
307             devFlags.enableGpuPmePPComm = false;
308         }
309     }
310
311     return devFlags;
312 }
313
314 /*! \brief Barrier for safe simultaneous thread access to mdrunner data
315  *
316  * Used to ensure that the master thread does not modify mdrunner during copy
317  * on the spawned threads. */
318 static void threadMpiMdrunnerAccessBarrier()
319 {
320 #if GMX_THREAD_MPI
321     MPI_Barrier(MPI_COMM_WORLD);
322 #endif
323 }
324
325 Mdrunner Mdrunner::cloneOnSpawnedThread() const
326 {
327     auto newRunner = Mdrunner(std::make_unique<MDModules>());
328
329     // All runners in the same process share a restraint manager resource because it is
330     // part of the interface to the client code, which is associated only with the
331     // original thread. Handles to the same resources can be obtained by copy.
332     {
333         newRunner.restraintManager_ = std::make_unique<RestraintManager>(*restraintManager_);
334     }
335
336     // Copy members of master runner.
337     // \todo Replace with builder when Simulation context and/or runner phases are better defined.
338     // Ref https://gitlab.com/gromacs/gromacs/-/issues/2587 and https://gitlab.com/gromacs/gromacs/-/issues/2375
339     newRunner.hw_opt    = hw_opt;
340     newRunner.filenames = filenames;
341
342     newRunner.oenv            = oenv;
343     newRunner.mdrunOptions    = mdrunOptions;
344     newRunner.domdecOptions   = domdecOptions;
345     newRunner.nbpu_opt        = nbpu_opt;
346     newRunner.pme_opt         = pme_opt;
347     newRunner.pme_fft_opt     = pme_fft_opt;
348     newRunner.bonded_opt      = bonded_opt;
349     newRunner.update_opt      = update_opt;
350     newRunner.nstlist_cmdline = nstlist_cmdline;
351     newRunner.replExParams    = replExParams;
352     newRunner.pforce          = pforce;
353     // Give the spawned thread the newly created valid communicator
354     // for the simulation.
355     newRunner.worldCommunicator   = MPI_COMM_WORLD;
356     newRunner.communicator        = MPI_COMM_WORLD;
357     newRunner.ms                  = ms;
358     newRunner.startingBehavior    = startingBehavior;
359     newRunner.stopHandlerBuilder_ = std::make_unique<StopHandlerBuilder>(*stopHandlerBuilder_);
360     newRunner.inputHolder_        = inputHolder_;
361
362     threadMpiMdrunnerAccessBarrier();
363
364     return newRunner;
365 }
366
367 /*! \brief The callback used for running on spawned threads.
368  *
369  * Obtains the pointer to the master mdrunner object from the one
370  * argument permitted to the thread-launch API call, copies it to make
371  * a new runner for this thread, reinitializes necessary data, and
372  * proceeds to the simulation. */
373 static void mdrunner_start_fn(const void* arg)
374 {
375     try
376     {
377         auto masterMdrunner = reinterpret_cast<const gmx::Mdrunner*>(arg);
378         /* copy the arg list to make sure that it's thread-local. This
379            doesn't copy pointed-to items, of course; fnm, cr and fplog
380            are reset in the call below, all others should be const. */
381         gmx::Mdrunner mdrunner = masterMdrunner->cloneOnSpawnedThread();
382         mdrunner.mdrunner();
383     }
384     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
385 }
386
387
388 void Mdrunner::spawnThreads(int numThreadsToLaunch)
389 {
390 #if GMX_THREAD_MPI
391     /* now spawn new threads that start mdrunner_start_fn(), while
392        the main thread returns. Thread affinity is handled later. */
393     if (tMPI_Init_fn(TRUE, numThreadsToLaunch, TMPI_AFFINITY_NONE, mdrunner_start_fn,
394                      static_cast<const void*>(this))
395         != TMPI_SUCCESS)
396     {
397         GMX_THROW(gmx::InternalError("Failed to spawn thread-MPI threads"));
398     }
399
400     // Give the master thread the newly created valid communicator for
401     // the simulation.
402     worldCommunicator = MPI_COMM_WORLD;
403     communicator      = MPI_COMM_WORLD;
404     threadMpiMdrunnerAccessBarrier();
405 #else
406     GMX_UNUSED_VALUE(numThreadsToLaunch);
407     GMX_UNUSED_VALUE(mdrunner_start_fn);
408 #endif
409 }
410
411 } // namespace gmx
412
413 /*! \brief Initialize variables for Verlet scheme simulation */
414 static void prepare_verlet_scheme(FILE*               fplog,
415                                   t_commrec*          cr,
416                                   t_inputrec*         ir,
417                                   int                 nstlist_cmdline,
418                                   const gmx_mtop_t*   mtop,
419                                   const matrix        box,
420                                   bool                makeGpuPairList,
421                                   const gmx::CpuInfo& cpuinfo)
422 {
423     // We checked the cut-offs in grompp, but double-check here.
424     // We have PME+LJcutoff kernels for rcoulomb>rvdw.
425     if (EEL_PME_EWALD(ir->coulombtype) && ir->vdwtype == eelCUT)
426     {
427         GMX_RELEASE_ASSERT(ir->rcoulomb >= ir->rvdw,
428                            "With Verlet lists and PME we should have rcoulomb>=rvdw");
429     }
430     else
431     {
432         GMX_RELEASE_ASSERT(ir->rcoulomb == ir->rvdw,
433                            "With Verlet lists and no PME rcoulomb and rvdw should be identical");
434     }
435     /* For NVE simulations, we will retain the initial list buffer */
436     if (EI_DYNAMICS(ir->eI) && ir->verletbuf_tol > 0 && !(EI_MD(ir->eI) && ir->etc == etcNO))
437     {
438         /* Update the Verlet buffer size for the current run setup */
439
440         /* Here we assume SIMD-enabled kernels are being used. But as currently
441          * calc_verlet_buffer_size gives the same results for 4x8 and 4x4
442          * and 4x2 gives a larger buffer than 4x4, this is ok.
443          */
444         ListSetupType listType =
445                 (makeGpuPairList ? ListSetupType::Gpu : ListSetupType::CpuSimdWhenSupported);
446         VerletbufListSetup listSetup = verletbufGetSafeListSetup(listType);
447
448         const real rlist_new =
449                 calcVerletBufferSize(*mtop, det(box), *ir, ir->nstlist, ir->nstlist - 1, -1, listSetup);
450
451         if (rlist_new != ir->rlist)
452         {
453             if (fplog != nullptr)
454             {
455                 fprintf(fplog,
456                         "\nChanging rlist from %g to %g for non-bonded %dx%d atom kernels\n\n",
457                         ir->rlist, rlist_new, listSetup.cluster_size_i, listSetup.cluster_size_j);
458             }
459             ir->rlist = rlist_new;
460         }
461     }
462
463     if (nstlist_cmdline > 0 && (!EI_DYNAMICS(ir->eI) || ir->verletbuf_tol <= 0))
464     {
465         gmx_fatal(FARGS, "Can not set nstlist without %s",
466                   !EI_DYNAMICS(ir->eI) ? "dynamics" : "verlet-buffer-tolerance");
467     }
468
469     if (EI_DYNAMICS(ir->eI))
470     {
471         /* Set or try nstlist values */
472         increaseNstlist(fplog, cr, ir, nstlist_cmdline, mtop, box, makeGpuPairList, cpuinfo);
473     }
474 }
475
476 /*! \brief Override the nslist value in inputrec
477  *
478  * with value passed on the command line (if any)
479  */
480 static void override_nsteps_cmdline(const gmx::MDLogger& mdlog, int64_t nsteps_cmdline, t_inputrec* ir)
481 {
482     assert(ir);
483
484     /* override with anything else than the default -2 */
485     if (nsteps_cmdline > -2)
486     {
487         char sbuf_steps[STEPSTRSIZE];
488         char sbuf_msg[STRLEN];
489
490         ir->nsteps = nsteps_cmdline;
491         if (EI_DYNAMICS(ir->eI) && nsteps_cmdline != -1)
492         {
493             sprintf(sbuf_msg,
494                     "Overriding nsteps with value passed on the command line: %s steps, %.3g ps",
495                     gmx_step_str(nsteps_cmdline, sbuf_steps), fabs(nsteps_cmdline * ir->delta_t));
496         }
497         else
498         {
499             sprintf(sbuf_msg, "Overriding nsteps with value passed on the command line: %s steps",
500                     gmx_step_str(nsteps_cmdline, sbuf_steps));
501         }
502
503         GMX_LOG(mdlog.warning).asParagraph().appendText(sbuf_msg);
504     }
505     else if (nsteps_cmdline < -2)
506     {
507         gmx_fatal(FARGS, "Invalid nsteps value passed on the command line: %" PRId64, nsteps_cmdline);
508     }
509     /* Do nothing if nsteps_cmdline == -2 */
510 }
511
512 namespace gmx
513 {
514
515 /*! \brief Return whether GPU acceleration of nonbondeds is supported with the given settings.
516  *
517  * If not, and if a warning may be issued, logs a warning about
518  * falling back to CPU code. With thread-MPI, only the first
519  * call to this function should have \c issueWarning true. */
520 static bool gpuAccelerationOfNonbondedIsUseful(const MDLogger& mdlog, const t_inputrec& ir, bool issueWarning)
521 {
522     bool        gpuIsUseful = true;
523     std::string warning;
524
525     if (ir.opts.ngener - ir.nwall > 1)
526     {
527         /* The GPU code does not support more than one energy group.
528          * If the user requested GPUs explicitly, a fatal error is given later.
529          */
530         gpuIsUseful = false;
531         warning =
532                 "Multiple energy groups is not implemented for GPUs, falling back to the CPU. "
533                 "For better performance, run on the GPU without energy groups and then do "
534                 "gmx mdrun -rerun option on the trajectory with an energy group .tpr file.";
535     }
536
537     if (EI_TPI(ir.eI))
538     {
539         gpuIsUseful = false;
540         warning     = "TPI is not implemented for GPUs.";
541     }
542
543     if (!gpuIsUseful && issueWarning)
544     {
545         GMX_LOG(mdlog.warning).asParagraph().appendText(warning);
546     }
547
548     return gpuIsUseful;
549 }
550
551 //! Initializes the logger for mdrun.
552 static gmx::LoggerOwner buildLogger(FILE* fplog, const bool isSimulationMasterRank)
553 {
554     gmx::LoggerBuilder builder;
555     if (fplog != nullptr)
556     {
557         builder.addTargetFile(gmx::MDLogger::LogLevel::Info, fplog);
558     }
559     if (isSimulationMasterRank)
560     {
561         builder.addTargetStream(gmx::MDLogger::LogLevel::Warning, &gmx::TextOutputFile::standardError());
562     }
563     return builder.build();
564 }
565
566 //! Make a TaskTarget from an mdrun argument string.
567 static TaskTarget findTaskTarget(const char* optionString)
568 {
569     TaskTarget returnValue = TaskTarget::Auto;
570
571     if (strncmp(optionString, "auto", 3) == 0)
572     {
573         returnValue = TaskTarget::Auto;
574     }
575     else if (strncmp(optionString, "cpu", 3) == 0)
576     {
577         returnValue = TaskTarget::Cpu;
578     }
579     else if (strncmp(optionString, "gpu", 3) == 0)
580     {
581         returnValue = TaskTarget::Gpu;
582     }
583     else
584     {
585         GMX_ASSERT(false, "Option string should have been checked for sanity already");
586     }
587
588     return returnValue;
589 }
590
591 //! Finish run, aggregate data to print performance info.
592 static void finish_run(FILE*                     fplog,
593                        const gmx::MDLogger&      mdlog,
594                        const t_commrec*          cr,
595                        const t_inputrec*         inputrec,
596                        t_nrnb                    nrnb[],
597                        gmx_wallcycle_t           wcycle,
598                        gmx_walltime_accounting_t walltime_accounting,
599                        nonbonded_verlet_t*       nbv,
600                        const gmx_pme_t*          pme,
601                        gmx_bool                  bWriteStat)
602 {
603     double delta_t = 0;
604     double nbfs = 0, mflop = 0;
605     double elapsed_time, elapsed_time_over_all_ranks, elapsed_time_over_all_threads,
606             elapsed_time_over_all_threads_over_all_ranks;
607     /* Control whether it is valid to print a report. Only the
608        simulation master may print, but it should not do so if the run
609        terminated e.g. before a scheduled reset step. This is
610        complicated by the fact that PME ranks are unaware of the
611        reason why they were sent a pmerecvqxFINISH. To avoid
612        communication deadlocks, we always do the communication for the
613        report, even if we've decided not to write the report, because
614        how long it takes to finish the run is not important when we've
615        decided not to report on the simulation performance.
616
617        Further, we only report performance for dynamical integrators,
618        because those are the only ones for which we plan to
619        consider doing any optimizations. */
620     bool printReport = EI_DYNAMICS(inputrec->eI) && SIMMASTER(cr);
621
622     if (printReport && !walltime_accounting_get_valid_finish(walltime_accounting))
623     {
624         GMX_LOG(mdlog.warning)
625                 .asParagraph()
626                 .appendText("Simulation ended prematurely, no performance report will be written.");
627         printReport = false;
628     }
629
630     t_nrnb*                 nrnb_tot;
631     std::unique_ptr<t_nrnb> nrnbTotalStorage;
632     if (cr->nnodes > 1)
633     {
634         nrnbTotalStorage = std::make_unique<t_nrnb>();
635         nrnb_tot         = nrnbTotalStorage.get();
636 #if GMX_MPI
637         MPI_Allreduce(nrnb->n, nrnb_tot->n, eNRNB, MPI_DOUBLE, MPI_SUM, cr->mpi_comm_mysim);
638 #endif
639     }
640     else
641     {
642         nrnb_tot = nrnb;
643     }
644
645     elapsed_time = walltime_accounting_get_time_since_reset(walltime_accounting);
646     elapsed_time_over_all_threads =
647             walltime_accounting_get_time_since_reset_over_all_threads(walltime_accounting);
648     if (cr->nnodes > 1)
649     {
650 #if GMX_MPI
651         /* reduce elapsed_time over all MPI ranks in the current simulation */
652         MPI_Allreduce(&elapsed_time, &elapsed_time_over_all_ranks, 1, MPI_DOUBLE, MPI_SUM,
653                       cr->mpi_comm_mysim);
654         elapsed_time_over_all_ranks /= cr->nnodes;
655         /* Reduce elapsed_time_over_all_threads over all MPI ranks in the
656          * current simulation. */
657         MPI_Allreduce(&elapsed_time_over_all_threads, &elapsed_time_over_all_threads_over_all_ranks,
658                       1, MPI_DOUBLE, MPI_SUM, cr->mpi_comm_mysim);
659 #endif
660     }
661     else
662     {
663         elapsed_time_over_all_ranks                  = elapsed_time;
664         elapsed_time_over_all_threads_over_all_ranks = elapsed_time_over_all_threads;
665     }
666
667     if (printReport)
668     {
669         print_flop(fplog, nrnb_tot, &nbfs, &mflop);
670     }
671
672     if (thisRankHasDuty(cr, DUTY_PP) && DOMAINDECOMP(cr))
673     {
674         print_dd_statistics(cr, inputrec, fplog);
675     }
676
677     /* TODO Move the responsibility for any scaling by thread counts
678      * to the code that handled the thread region, so that there's a
679      * mechanism to keep cycle counting working during the transition
680      * to task parallelism. */
681     int nthreads_pp  = gmx_omp_nthreads_get(emntNonbonded);
682     int nthreads_pme = gmx_omp_nthreads_get(emntPME);
683     wallcycle_scale_by_num_threads(wcycle, thisRankHasDuty(cr, DUTY_PME) && !thisRankHasDuty(cr, DUTY_PP),
684                                    nthreads_pp, nthreads_pme);
685     auto cycle_sum(wallcycle_sum(cr, wcycle));
686
687     if (printReport)
688     {
689         auto nbnxn_gpu_timings =
690                 (nbv != nullptr && nbv->useGpu()) ? Nbnxm::gpu_get_timings(nbv->gpu_nbv) : nullptr;
691         gmx_wallclock_gpu_pme_t pme_gpu_timings = {};
692
693         if (pme_gpu_task_enabled(pme))
694         {
695             pme_gpu_get_timings(pme, &pme_gpu_timings);
696         }
697         wallcycle_print(fplog, mdlog, cr->nnodes, cr->npmenodes, nthreads_pp, nthreads_pme,
698                         elapsed_time_over_all_ranks, wcycle, cycle_sum, nbnxn_gpu_timings,
699                         &pme_gpu_timings);
700
701         if (EI_DYNAMICS(inputrec->eI))
702         {
703             delta_t = inputrec->delta_t;
704         }
705
706         if (fplog)
707         {
708             print_perf(fplog, elapsed_time_over_all_threads_over_all_ranks, elapsed_time_over_all_ranks,
709                        walltime_accounting_get_nsteps_done_since_reset(walltime_accounting),
710                        delta_t, nbfs, mflop);
711         }
712         if (bWriteStat)
713         {
714             print_perf(stderr, elapsed_time_over_all_threads_over_all_ranks, elapsed_time_over_all_ranks,
715                        walltime_accounting_get_nsteps_done_since_reset(walltime_accounting),
716                        delta_t, nbfs, mflop);
717         }
718     }
719 }
720
721 int Mdrunner::mdrunner()
722 {
723     matrix                    box;
724     t_forcerec*               fr               = nullptr;
725     real                      ewaldcoeff_q     = 0;
726     real                      ewaldcoeff_lj    = 0;
727     int                       nChargePerturbed = -1, nTypePerturbed = 0;
728     gmx_wallcycle_t           wcycle;
729     gmx_walltime_accounting_t walltime_accounting = nullptr;
730     MembedHolder              membedHolder(filenames.size(), filenames.data());
731     gmx_hw_info_t*            hwinfo = nullptr;
732
733     /* CAUTION: threads may be started later on in this function, so
734        cr doesn't reflect the final parallel state right now */
735     gmx_mtop_t mtop;
736
737     /* TODO: inputrec should tell us whether we use an algorithm, not a file option */
738     const bool doEssentialDynamics = opt2bSet("-ei", filenames.size(), filenames.data());
739     const bool doRerun             = mdrunOptions.rerun;
740
741     // Handle task-assignment related user options.
742     EmulateGpuNonbonded emulateGpuNonbonded =
743             (getenv("GMX_EMULATE_GPU") != nullptr ? EmulateGpuNonbonded::Yes : EmulateGpuNonbonded::No);
744
745     std::vector<int> userGpuTaskAssignment;
746     try
747     {
748         userGpuTaskAssignment = parseUserTaskAssignmentString(hw_opt.userGpuTaskAssignment);
749     }
750     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
751     auto nonbondedTarget = findTaskTarget(nbpu_opt);
752     auto pmeTarget       = findTaskTarget(pme_opt);
753     auto pmeFftTarget    = findTaskTarget(pme_fft_opt);
754     auto bondedTarget    = findTaskTarget(bonded_opt);
755     auto updateTarget    = findTaskTarget(update_opt);
756
757     FILE* fplog = nullptr;
758     // If we are appending, we don't write log output because we need
759     // to check that the old log file matches what the checkpoint file
760     // expects. Otherwise, we should start to write log output now if
761     // there is a file ready for it.
762     if (logFileHandle != nullptr && startingBehavior != StartingBehavior::RestartWithAppending)
763     {
764         fplog = gmx_fio_getfp(logFileHandle);
765     }
766     const bool       isSimulationMasterRank = findIsSimulationMasterRank(ms, communicator);
767     gmx::LoggerOwner logOwner(buildLogger(fplog, isSimulationMasterRank));
768     gmx::MDLogger    mdlog(logOwner.logger());
769
770     // TODO The thread-MPI master rank makes a working
771     // PhysicalNodeCommunicator here, but it gets rebuilt by all ranks
772     // after the threads have been launched. This works because no use
773     // is made of that communicator until after the execution paths
774     // have rejoined. But it is likely that we can improve the way
775     // this is expressed, e.g. by expressly running detection only the
776     // master rank for thread-MPI, rather than relying on the mutex
777     // and reference count.
778     PhysicalNodeCommunicator physicalNodeComm(worldCommunicator, gmx_physicalnode_id_hash());
779     hwinfo = gmx_detect_hardware(mdlog, physicalNodeComm);
780
781     gmx_print_detected_hardware(fplog, isSimulationMasterRank && isMasterSim(ms), mdlog, hwinfo);
782
783     std::vector<int> gpuIdsToUse = makeGpuIdsToUse(hwinfo->deviceInfoList, hw_opt.gpuIdsAvailable);
784
785     // Print citation requests after all software/hardware printing
786     pleaseCiteGromacs(fplog);
787
788     // Note: legacy program logic relies on checking whether these pointers are assigned.
789     // Objects may or may not be allocated later.
790     std::unique_ptr<t_inputrec> inputrec;
791     std::unique_ptr<t_state>    globalState;
792
793     auto partialDeserializedTpr = std::make_unique<PartialDeserializedTprFile>();
794
795     if (isSimulationMasterRank)
796     {
797         // Allocate objects to be initialized by later function calls.
798         /* Only the master rank has the global state */
799         globalState = std::make_unique<t_state>();
800         inputrec    = std::make_unique<t_inputrec>();
801
802         /* Read (nearly) all data required for the simulation
803          * and keep the partly serialized tpr contents to send to other ranks later
804          */
805         applyGlobalSimulationState(*inputHolder_.get(), partialDeserializedTpr.get(),
806                                    globalState.get(), inputrec.get(), &mtop);
807     }
808
809     /* Check and update the hardware options for internal consistency */
810     checkAndUpdateHardwareOptions(mdlog, &hw_opt, isSimulationMasterRank, domdecOptions.numPmeRanks,
811                                   inputrec.get());
812
813     if (GMX_THREAD_MPI && isSimulationMasterRank)
814     {
815         bool useGpuForNonbonded = false;
816         bool useGpuForPme       = false;
817         try
818         {
819             GMX_RELEASE_ASSERT(inputrec != nullptr, "Keep the compiler happy");
820
821             // If the user specified the number of ranks, then we must
822             // respect that, but in default mode, we need to allow for
823             // the number of GPUs to choose the number of ranks.
824             auto canUseGpuForNonbonded = buildSupportsNonbondedOnGpu(nullptr);
825             useGpuForNonbonded         = decideWhetherToUseGpusForNonbondedWithThreadMpi(
826                     nonbondedTarget, gpuIdsToUse, userGpuTaskAssignment, emulateGpuNonbonded,
827                     canUseGpuForNonbonded,
828                     gpuAccelerationOfNonbondedIsUseful(mdlog, *inputrec, GMX_THREAD_MPI),
829                     hw_opt.nthreads_tmpi);
830             useGpuForPme = decideWhetherToUseGpusForPmeWithThreadMpi(
831                     useGpuForNonbonded, pmeTarget, gpuIdsToUse, userGpuTaskAssignment, *hwinfo,
832                     *inputrec, hw_opt.nthreads_tmpi, domdecOptions.numPmeRanks);
833         }
834         GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
835
836         /* Determine how many thread-MPI ranks to start.
837          *
838          * TODO Over-writing the user-supplied value here does
839          * prevent any possible subsequent checks from working
840          * correctly. */
841         hw_opt.nthreads_tmpi =
842                 get_nthreads_mpi(hwinfo, &hw_opt, gpuIdsToUse, useGpuForNonbonded, useGpuForPme,
843                                  inputrec.get(), &mtop, mdlog, membedHolder.doMembed());
844
845         // Now start the threads for thread MPI.
846         spawnThreads(hw_opt.nthreads_tmpi);
847         // The spawned threads enter mdrunner() and execution of
848         // master and spawned threads joins at the end of this block.
849         physicalNodeComm = PhysicalNodeCommunicator(worldCommunicator, gmx_physicalnode_id_hash());
850     }
851
852     GMX_RELEASE_ASSERT(ms || communicator == MPI_COMM_WORLD,
853                        "Must have valid world communicator unless running a multi-simulation");
854     CommrecHandle crHandle = init_commrec(communicator);
855     t_commrec*    cr       = crHandle.get();
856     GMX_RELEASE_ASSERT(cr != nullptr, "Must have valid commrec");
857
858     if (PAR(cr))
859     {
860         /* now broadcast everything to the non-master nodes/threads: */
861         if (!isSimulationMasterRank)
862         {
863             // Until now, only the master rank has a non-null pointer.
864             // On non-master ranks, allocate the object that will receive data in the following call.
865             inputrec = std::make_unique<t_inputrec>();
866         }
867         init_parallel(cr->mpiDefaultCommunicator, MASTER(cr), inputrec.get(), &mtop,
868                       partialDeserializedTpr.get());
869     }
870     GMX_RELEASE_ASSERT(inputrec != nullptr, "All ranks should have a valid inputrec now");
871     partialDeserializedTpr.reset(nullptr);
872
873     // Now the number of ranks is known to all ranks, and each knows
874     // the inputrec read by the master rank. The ranks can now all run
875     // the task-deciding functions and will agree on the result
876     // without needing to communicate.
877     const bool useDomainDecomposition = (PAR(cr) && !(EI_TPI(inputrec->eI) || inputrec->eI == eiNM));
878
879     // Note that these variables describe only their own node.
880     //
881     // Note that when bonded interactions run on a GPU they always run
882     // alongside a nonbonded task, so do not influence task assignment
883     // even though they affect the force calculation workload.
884     bool useGpuForNonbonded = false;
885     bool useGpuForPme       = false;
886     bool useGpuForBonded    = false;
887     bool useGpuForUpdate    = false;
888     bool gpusWereDetected   = hwinfo->ngpu_compatible_tot > 0;
889     try
890     {
891         // It's possible that there are different numbers of GPUs on
892         // different nodes, which is the user's responsibility to
893         // handle. If unsuitable, we will notice that during task
894         // assignment.
895         auto canUseGpuForNonbonded = buildSupportsNonbondedOnGpu(nullptr);
896         useGpuForNonbonded         = decideWhetherToUseGpusForNonbonded(
897                 nonbondedTarget, userGpuTaskAssignment, emulateGpuNonbonded, canUseGpuForNonbonded,
898                 gpuAccelerationOfNonbondedIsUseful(mdlog, *inputrec, !GMX_THREAD_MPI), gpusWereDetected);
899         useGpuForPme = decideWhetherToUseGpusForPme(
900                 useGpuForNonbonded, pmeTarget, userGpuTaskAssignment, *hwinfo, *inputrec,
901                 cr->sizeOfDefaultCommunicator, domdecOptions.numPmeRanks, gpusWereDetected);
902         auto canUseGpuForBonded = buildSupportsGpuBondeds(nullptr)
903                                   && inputSupportsGpuBondeds(*inputrec, mtop, nullptr);
904         useGpuForBonded = decideWhetherToUseGpusForBonded(
905                 useGpuForNonbonded, useGpuForPme, bondedTarget, canUseGpuForBonded,
906                 EVDW_PME(inputrec->vdwtype), EEL_PME_EWALD(inputrec->coulombtype),
907                 domdecOptions.numPmeRanks, gpusWereDetected);
908     }
909     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
910
911     const PmeRunMode pmeRunMode = determinePmeRunMode(useGpuForPme, pmeFftTarget, *inputrec);
912
913     // Initialize development feature flags that enabled by environment variable
914     // and report those features that are enabled.
915     const DevelopmentFeatureFlags devFlags =
916             manageDevelopmentFeatures(mdlog, useGpuForNonbonded, pmeRunMode);
917
918     const bool useModularSimulator =
919             checkUseModularSimulator(false, inputrec.get(), doRerun, mtop, ms, replExParams,
920                                      nullptr, doEssentialDynamics, membedHolder.doMembed());
921
922     // Build restraints.
923     // TODO: hide restraint implementation details from Mdrunner.
924     // There is nothing unique about restraints at this point as far as the
925     // Mdrunner is concerned. The Mdrunner should just be getting a sequence of
926     // factory functions from the SimulationContext on which to call mdModules_->add().
927     // TODO: capture all restraints into a single RestraintModule, passed to the runner builder.
928     for (auto&& restraint : restraintManager_->getRestraints())
929     {
930         auto module = RestraintMDModule::create(restraint, restraint->sites());
931         mdModules_->add(std::move(module));
932     }
933
934     // TODO: Error handling
935     mdModules_->assignOptionsToModules(*inputrec->params, nullptr);
936     // now that the MdModules know their options, they know which callbacks to sign up to
937     mdModules_->subscribeToSimulationSetupNotifications();
938     const auto& mdModulesNotifier = mdModules_->notifier().simulationSetupNotifications_;
939
940     if (inputrec->internalParameters != nullptr)
941     {
942         mdModulesNotifier.notify(*inputrec->internalParameters);
943     }
944
945     if (fplog != nullptr)
946     {
947         pr_inputrec(fplog, 0, "Input Parameters", inputrec.get(), FALSE);
948         fprintf(fplog, "\n");
949     }
950
951     if (SIMMASTER(cr))
952     {
953         /* In rerun, set velocities to zero if present */
954         if (doRerun && ((globalState->flags & (1 << estV)) != 0))
955         {
956             // rerun does not use velocities
957             GMX_LOG(mdlog.info)
958                     .asParagraph()
959                     .appendText(
960                             "Rerun trajectory contains velocities. Rerun does only evaluate "
961                             "potential energy and forces. The velocities will be ignored.");
962             for (int i = 0; i < globalState->natoms; i++)
963             {
964                 clear_rvec(globalState->v[i]);
965             }
966             globalState->flags &= ~(1 << estV);
967         }
968
969         /* now make sure the state is initialized and propagated */
970         set_state_entries(globalState.get(), inputrec.get(), useModularSimulator);
971     }
972
973     /* NM and TPI parallelize over force/energy calculations, not atoms,
974      * so we need to initialize and broadcast the global state.
975      */
976     if (inputrec->eI == eiNM || inputrec->eI == eiTPI)
977     {
978         if (!MASTER(cr))
979         {
980             globalState = std::make_unique<t_state>();
981         }
982         broadcastStateWithoutDynamics(cr->mpiDefaultCommunicator, DOMAINDECOMP(cr), PAR(cr),
983                                       globalState.get());
984     }
985
986     /* A parallel command line option consistency check that we can
987        only do after any threads have started. */
988     if (!PAR(cr)
989         && (domdecOptions.numCells[XX] > 1 || domdecOptions.numCells[YY] > 1
990             || domdecOptions.numCells[ZZ] > 1 || domdecOptions.numPmeRanks > 0))
991     {
992         gmx_fatal(FARGS,
993                   "The -dd or -npme option request a parallel simulation, "
994 #if !GMX_MPI
995                   "but %s was compiled without threads or MPI enabled",
996                   output_env_get_program_display_name(oenv));
997 #elif GMX_THREAD_MPI
998                   "but the number of MPI-threads (option -ntmpi) is not set or is 1");
999 #else
1000                   "but %s was not started through mpirun/mpiexec or only one rank was requested "
1001                   "through mpirun/mpiexec",
1002                   output_env_get_program_display_name(oenv));
1003 #endif
1004     }
1005
1006     if (doRerun && (EI_ENERGY_MINIMIZATION(inputrec->eI) || eiNM == inputrec->eI))
1007     {
1008         gmx_fatal(FARGS,
1009                   "The .mdp file specified an energy mininization or normal mode algorithm, and "
1010                   "these are not compatible with mdrun -rerun");
1011     }
1012
1013     if (!(EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype)))
1014     {
1015         if (domdecOptions.numPmeRanks > 0)
1016         {
1017             gmx_fatal_collective(FARGS, cr->mpiDefaultCommunicator, MASTER(cr),
1018                                  "PME-only ranks are requested, but the system does not use PME "
1019                                  "for electrostatics or LJ");
1020         }
1021
1022         domdecOptions.numPmeRanks = 0;
1023     }
1024
1025     if (useGpuForNonbonded && domdecOptions.numPmeRanks < 0)
1026     {
1027         /* With NB GPUs we don't automatically use PME-only CPU ranks. PME ranks can
1028          * improve performance with many threads per GPU, since our OpenMP
1029          * scaling is bad, but it's difficult to automate the setup.
1030          */
1031         domdecOptions.numPmeRanks = 0;
1032     }
1033     if (useGpuForPme)
1034     {
1035         if (domdecOptions.numPmeRanks < 0)
1036         {
1037             domdecOptions.numPmeRanks = 0;
1038             // TODO possibly print a note that one can opt-in for a separate PME GPU rank?
1039         }
1040         else
1041         {
1042             GMX_RELEASE_ASSERT(domdecOptions.numPmeRanks <= 1,
1043                                "PME GPU decomposition is not supported");
1044         }
1045     }
1046
1047 #if GMX_FAHCORE
1048     if (MASTER(cr))
1049     {
1050         fcRegisterSteps(inputrec->nsteps, inputrec->init_step);
1051     }
1052 #endif
1053
1054     /* NMR restraints must be initialized before load_checkpoint,
1055      * since with time averaging the history is added to t_state.
1056      * For proper consistency check we therefore need to extend
1057      * t_state here.
1058      * So the PME-only nodes (if present) will also initialize
1059      * the distance restraints.
1060      */
1061
1062     /* This needs to be called before read_checkpoint to extend the state */
1063     t_disresdata* disresdata;
1064     snew(disresdata, 1);
1065     init_disres(fplog, &mtop, inputrec.get(), DisResRunMode::MDRun,
1066                 MASTER(cr) ? DDRole::Master : DDRole::Agent,
1067                 PAR(cr) ? NumRanks::Multiple : NumRanks::Single, cr->mpi_comm_mysim, ms, disresdata,
1068                 globalState.get(), replExParams.exchangeInterval > 0);
1069
1070     t_oriresdata* oriresdata;
1071     snew(oriresdata, 1);
1072     init_orires(fplog, &mtop, inputrec.get(), cr, ms, globalState.get(), oriresdata);
1073
1074     auto deform = prepareBoxDeformation(
1075             globalState != nullptr ? globalState->box : box, MASTER(cr) ? DDRole::Master : DDRole::Agent,
1076             PAR(cr) ? NumRanks::Multiple : NumRanks::Single, cr->mpi_comm_mygroup, *inputrec);
1077
1078     ObservablesHistory observablesHistory = {};
1079
1080     auto modularSimulatorCheckpointData = std::make_unique<ReadCheckpointDataHolder>();
1081     if (startingBehavior != StartingBehavior::NewSimulation)
1082     {
1083         /* Check if checkpoint file exists before doing continuation.
1084          * This way we can use identical input options for the first and subsequent runs...
1085          */
1086         if (mdrunOptions.numStepsCommandline > -2)
1087         {
1088             /* Temporarily set the number of steps to unlimited to avoid
1089              * triggering the nsteps check in load_checkpoint().
1090              * This hack will go away soon when the -nsteps option is removed.
1091              */
1092             inputrec->nsteps = -1;
1093         }
1094
1095         // Finish applying initial simulation state information from external sources on all ranks.
1096         // Reconcile checkpoint file data with Mdrunner state established up to this point.
1097         applyLocalState(*inputHolder_.get(), logFileHandle, cr, domdecOptions.numCells,
1098                         inputrec.get(), globalState.get(), &observablesHistory,
1099                         mdrunOptions.reproducible, mdModules_->notifier(),
1100                         modularSimulatorCheckpointData.get(), useModularSimulator);
1101         // TODO: (#3652) Synchronize filesystem state, SimulationInput contents, and program
1102         //  invariants
1103         //  on all code paths.
1104         // Write checkpoint or provide hook to update SimulationInput.
1105         // If there was a checkpoint file, SimulationInput contains more information
1106         // than if there wasn't. At this point, we have synchronized the in-memory
1107         // state with the filesystem state only for restarted simulations. We should
1108         // be calling applyLocalState unconditionally and expect that the completeness
1109         // of SimulationInput is not dependent on its creation method.
1110
1111         if (startingBehavior == StartingBehavior::RestartWithAppending && logFileHandle)
1112         {
1113             // Now we can start normal logging to the truncated log file.
1114             fplog = gmx_fio_getfp(logFileHandle);
1115             prepareLogAppending(fplog);
1116             logOwner = buildLogger(fplog, MASTER(cr));
1117             mdlog    = logOwner.logger();
1118         }
1119     }
1120
1121     if (mdrunOptions.numStepsCommandline > -2)
1122     {
1123         GMX_LOG(mdlog.info)
1124                 .asParagraph()
1125                 .appendText(
1126                         "The -nsteps functionality is deprecated, and may be removed in a future "
1127                         "version. "
1128                         "Consider using gmx convert-tpr -nsteps or changing the appropriate .mdp "
1129                         "file field.");
1130     }
1131     /* override nsteps with value set on the commandline */
1132     override_nsteps_cmdline(mdlog, mdrunOptions.numStepsCommandline, inputrec.get());
1133
1134     if (isSimulationMasterRank)
1135     {
1136         copy_mat(globalState->box, box);
1137     }
1138
1139     if (PAR(cr))
1140     {
1141         gmx_bcast(sizeof(box), box, cr->mpiDefaultCommunicator);
1142     }
1143
1144     if (inputrec->cutoff_scheme != ecutsVERLET)
1145     {
1146         gmx_fatal(FARGS,
1147                   "This group-scheme .tpr file can no longer be run by mdrun. Please update to the "
1148                   "Verlet scheme, or use an earlier version of GROMACS if necessary.");
1149     }
1150     /* Update rlist and nstlist. */
1151     /* Note: prepare_verlet_scheme is calling increaseNstlist(...), which (while attempting to
1152      * increase rlist) tries to check if the newly chosen value fits with the DD scheme. As this is
1153      * run before any DD scheme is set up, this check is never executed. See #3334 for more details.
1154      */
1155     prepare_verlet_scheme(fplog, cr, inputrec.get(), nstlist_cmdline, &mtop, box,
1156                           useGpuForNonbonded || (emulateGpuNonbonded == EmulateGpuNonbonded::Yes),
1157                           *hwinfo->cpuInfo);
1158
1159     const bool prefer1DAnd1PulseDD = (devFlags.enableGpuHaloExchange && useGpuForNonbonded);
1160     // This builder is necessary while we have multi-part construction
1161     // of DD. Before DD is constructed, we use the existence of
1162     // the builder object to indicate that further construction of DD
1163     // is needed.
1164     std::unique_ptr<DomainDecompositionBuilder> ddBuilder;
1165     if (useDomainDecomposition)
1166     {
1167         ddBuilder = std::make_unique<DomainDecompositionBuilder>(
1168                 mdlog, cr, domdecOptions, mdrunOptions, prefer1DAnd1PulseDD, mtop, *inputrec, box,
1169                 positionsFromStatePointer(globalState.get()));
1170     }
1171     else
1172     {
1173         /* PME, if used, is done on all nodes with 1D decomposition */
1174         cr->nnodes     = cr->sizeOfDefaultCommunicator;
1175         cr->sim_nodeid = cr->rankInDefaultCommunicator;
1176         cr->nodeid     = cr->rankInDefaultCommunicator;
1177         cr->npmenodes  = 0;
1178         cr->duty       = (DUTY_PP | DUTY_PME);
1179
1180         if (inputrec->pbcType == PbcType::Screw)
1181         {
1182             gmx_fatal(FARGS, "pbc=screw is only implemented with domain decomposition");
1183         }
1184     }
1185
1186     // Produce the task assignment for this rank - done after DD is constructed
1187     GpuTaskAssignments gpuTaskAssignments = GpuTaskAssignmentsBuilder::build(
1188             gpuIdsToUse, userGpuTaskAssignment, *hwinfo, communicator, physicalNodeComm,
1189             nonbondedTarget, pmeTarget, bondedTarget, updateTarget, useGpuForNonbonded,
1190             useGpuForPme, thisRankHasDuty(cr, DUTY_PP),
1191             // TODO cr->duty & DUTY_PME should imply that a PME
1192             // algorithm is active, but currently does not.
1193             EEL_PME(inputrec->coulombtype) && thisRankHasDuty(cr, DUTY_PME));
1194
1195     // Get the device handles for the modules, nullptr when no task is assigned.
1196     int                deviceId   = -1;
1197     DeviceInformation* deviceInfo = gpuTaskAssignments.initDevice(&deviceId);
1198
1199     // timing enabling - TODO put this in gpu_utils (even though generally this is just option handling?)
1200     bool useTiming = true;
1201
1202     if (GMX_GPU_CUDA)
1203     {
1204         /* WARNING: CUDA timings are incorrect with multiple streams.
1205          *          This is the main reason why they are disabled by default.
1206          */
1207         // TODO: Consider turning on by default when we can detect nr of streams.
1208         useTiming = (getenv("GMX_ENABLE_GPU_TIMING") != nullptr);
1209     }
1210     else if (GMX_GPU_OPENCL)
1211     {
1212         useTiming = (getenv("GMX_DISABLE_GPU_TIMING") == nullptr);
1213     }
1214
1215     // TODO Currently this is always built, yet DD partition code
1216     // checks if it is built before using it. Probably it should
1217     // become an MDModule that is made only when another module
1218     // requires it (e.g. pull, CompEl, density fitting), so that we
1219     // don't update the local atom sets unilaterally every step.
1220     LocalAtomSetManager atomSets;
1221     if (ddBuilder)
1222     {
1223         // TODO Pass the GPU streams to ddBuilder to use in buffer
1224         // transfers (e.g. halo exchange)
1225         cr->dd = ddBuilder->build(&atomSets);
1226         // The builder's job is done, so destruct it
1227         ddBuilder.reset(nullptr);
1228         // Note that local state still does not exist yet.
1229     }
1230
1231     // The GPU update is decided here because we need to know whether the constraints or
1232     // SETTLEs can span accross the domain borders (i.e. whether or not update groups are
1233     // defined). This is only known after DD is initialized, hence decision on using GPU
1234     // update is done so late.
1235     try
1236     {
1237         const bool useUpdateGroups = cr->dd ? ddUsesUpdateGroups(*cr->dd) : false;
1238
1239         useGpuForUpdate = decideWhetherToUseGpuForUpdate(
1240                 useDomainDecomposition, useUpdateGroups, pmeRunMode, domdecOptions.numPmeRanks > 0,
1241                 useGpuForNonbonded, updateTarget, gpusWereDetected, *inputrec, mtop,
1242                 doEssentialDynamics, gmx_mtop_ftype_count(mtop, F_ORIRES) > 0,
1243                 replExParams.exchangeInterval > 0, doRerun, devFlags, mdlog);
1244     }
1245     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1246
1247     const bool printHostName = (cr->nnodes > 1);
1248     gpuTaskAssignments.reportGpuUsage(mdlog, printHostName, useGpuForBonded, pmeRunMode, useGpuForUpdate);
1249
1250     MdrunScheduleWorkload runScheduleWork;
1251     // Also populates the simulation constant workload description.
1252     runScheduleWork.simulationWork = createSimulationWorkload(
1253             *inputrec, devFlags, useGpuForNonbonded, pmeRunMode, useGpuForBonded, useGpuForUpdate);
1254
1255     std::unique_ptr<DeviceStreamManager> deviceStreamManager = nullptr;
1256
1257     if (deviceInfo != nullptr)
1258     {
1259         if (DOMAINDECOMP(cr) && thisRankHasDuty(cr, DUTY_PP))
1260         {
1261             dd_setup_dlb_resource_sharing(cr, deviceId);
1262         }
1263         deviceStreamManager = std::make_unique<DeviceStreamManager>(
1264                 *deviceInfo, havePPDomainDecomposition(cr), runScheduleWork.simulationWork, useTiming);
1265     }
1266
1267     // If the user chose a task assignment, give them some hints
1268     // where appropriate.
1269     if (!userGpuTaskAssignment.empty())
1270     {
1271         gpuTaskAssignments.logPerformanceHints(mdlog, ssize(gpuIdsToUse));
1272     }
1273
1274     if (PAR(cr))
1275     {
1276         /* After possible communicator splitting in make_dd_communicators.
1277          * we can set up the intra/inter node communication.
1278          */
1279         gmx_setup_nodecomm(fplog, cr);
1280     }
1281
1282 #if GMX_MPI
1283     if (isMultiSim(ms))
1284     {
1285         GMX_LOG(mdlog.warning)
1286                 .asParagraph()
1287                 .appendTextFormatted(
1288                         "This is simulation %d out of %d running as a composite GROMACS\n"
1289                         "multi-simulation job. Setup for this simulation:\n",
1290                         ms->simulationIndex_, ms->numSimulations_);
1291     }
1292     GMX_LOG(mdlog.warning)
1293             .appendTextFormatted("Using %d MPI %s\n", cr->nnodes,
1294 #    if GMX_THREAD_MPI
1295                                  cr->nnodes == 1 ? "thread" : "threads"
1296 #    else
1297                                  cr->nnodes == 1 ? "process" : "processes"
1298 #    endif
1299             );
1300     fflush(stderr);
1301 #endif
1302
1303     // If mdrun -pin auto honors any affinity setting that already
1304     // exists. If so, it is nice to provide feedback about whether
1305     // that existing affinity setting was from OpenMP or something
1306     // else, so we run this code both before and after we initialize
1307     // the OpenMP support.
1308     gmx_check_thread_affinity_set(mdlog, &hw_opt, hwinfo->nthreads_hw_avail, FALSE);
1309     /* Check and update the number of OpenMP threads requested */
1310     checkAndUpdateRequestedNumOpenmpThreads(&hw_opt, *hwinfo, cr, ms, physicalNodeComm.size_,
1311                                             pmeRunMode, mtop, *inputrec);
1312
1313     gmx_omp_nthreads_init(mdlog, cr, hwinfo->nthreads_hw_avail, physicalNodeComm.size_,
1314                           hw_opt.nthreads_omp, hw_opt.nthreads_omp_pme, !thisRankHasDuty(cr, DUTY_PP));
1315
1316     // Enable FP exception detection, but not in
1317     // Release mode and not for compilers with known buggy FP
1318     // exception support (clang with any optimization) or suspected
1319     // buggy FP exception support (gcc 7.* with optimization).
1320 #if !defined NDEBUG                                                                         \
1321         && !((defined __clang__ || (defined(__GNUC__) && !defined(__ICC) && __GNUC__ == 7)) \
1322              && defined __OPTIMIZE__)
1323     const bool bEnableFPE = true;
1324 #else
1325     const bool bEnableFPE = false;
1326 #endif
1327     // FIXME - reconcile with gmx_feenableexcept() call from CommandLineModuleManager::run()
1328     if (bEnableFPE)
1329     {
1330         gmx_feenableexcept();
1331     }
1332
1333     /* Now that we know the setup is consistent, check for efficiency */
1334     check_resource_division_efficiency(hwinfo, gpuTaskAssignments.thisRankHasAnyGpuTask(),
1335                                        mdrunOptions.ntompOptionIsSet, cr, mdlog);
1336
1337     /* getting number of PP/PME threads on this MPI / tMPI rank.
1338        PME: env variable should be read only on one node to make sure it is
1339        identical everywhere;
1340      */
1341     const int numThreadsOnThisRank = thisRankHasDuty(cr, DUTY_PP) ? gmx_omp_nthreads_get(emntNonbonded)
1342                                                                   : gmx_omp_nthreads_get(emntPME);
1343     checkHardwareOversubscription(numThreadsOnThisRank, cr->nodeid, *hwinfo->hardwareTopology,
1344                                   physicalNodeComm, mdlog);
1345
1346     // Enable Peer access between GPUs where available
1347     // Only for DD, only master PP rank needs to perform setup, and only if thread MPI plus
1348     // any of the GPU communication features are active.
1349     if (DOMAINDECOMP(cr) && MASTER(cr) && thisRankHasDuty(cr, DUTY_PP) && GMX_THREAD_MPI
1350         && (runScheduleWork.simulationWork.useGpuHaloExchange
1351             || runScheduleWork.simulationWork.useGpuPmePpCommunication))
1352     {
1353         setupGpuDevicePeerAccess(gpuIdsToUse, mdlog);
1354     }
1355
1356     if (hw_opt.threadAffinity != ThreadAffinity::Off)
1357     {
1358         /* Before setting affinity, check whether the affinity has changed
1359          * - which indicates that probably the OpenMP library has changed it
1360          * since we first checked).
1361          */
1362         gmx_check_thread_affinity_set(mdlog, &hw_opt, hwinfo->nthreads_hw_avail, TRUE);
1363
1364         int numThreadsOnThisNode, intraNodeThreadOffset;
1365         analyzeThreadsOnThisNode(physicalNodeComm, numThreadsOnThisRank, &numThreadsOnThisNode,
1366                                  &intraNodeThreadOffset);
1367
1368         /* Set the CPU affinity */
1369         gmx_set_thread_affinity(mdlog, cr, &hw_opt, *hwinfo->hardwareTopology, numThreadsOnThisRank,
1370                                 numThreadsOnThisNode, intraNodeThreadOffset, nullptr);
1371     }
1372
1373     if (mdrunOptions.timingOptions.resetStep > -1)
1374     {
1375         GMX_LOG(mdlog.info)
1376                 .asParagraph()
1377                 .appendText(
1378                         "The -resetstep functionality is deprecated, and may be removed in a "
1379                         "future version.");
1380     }
1381     wcycle = wallcycle_init(fplog, mdrunOptions.timingOptions.resetStep, cr);
1382
1383     if (PAR(cr))
1384     {
1385         /* Master synchronizes its value of reset_counters with all nodes
1386          * including PME only nodes */
1387         int64_t reset_counters = wcycle_get_reset_counters(wcycle);
1388         gmx_bcast(sizeof(reset_counters), &reset_counters, cr->mpi_comm_mysim);
1389         wcycle_set_reset_counters(wcycle, reset_counters);
1390     }
1391
1392     // Membrane embedding must be initialized before we call init_forcerec()
1393     membedHolder.initializeMembed(fplog, filenames.size(), filenames.data(), &mtop, inputrec.get(),
1394                                   globalState.get(), cr, &mdrunOptions.checkpointOptions.period);
1395
1396     const bool               thisRankHasPmeGpuTask = gpuTaskAssignments.thisRankHasPmeGpuTask();
1397     std::unique_ptr<MDAtoms> mdAtoms;
1398     std::unique_ptr<VirtualSitesHandler> vsite;
1399     std::unique_ptr<GpuBonded>           gpuBonded;
1400
1401     t_nrnb nrnb;
1402     if (thisRankHasDuty(cr, DUTY_PP))
1403     {
1404         mdModulesNotifier.notify(*cr);
1405         mdModulesNotifier.notify(&atomSets);
1406         mdModulesNotifier.notify(inputrec->pbcType);
1407         mdModulesNotifier.notify(SimulationTimeStep{ inputrec->delta_t });
1408         /* Initiate forcerecord */
1409         fr                 = new t_forcerec;
1410         fr->forceProviders = mdModules_->initForceProviders();
1411         init_forcerec(fplog, mdlog, fr, inputrec.get(), &mtop, cr, box,
1412                       opt2fn("-table", filenames.size(), filenames.data()),
1413                       opt2fn("-tablep", filenames.size(), filenames.data()),
1414                       opt2fns("-tableb", filenames.size(), filenames.data()), pforce);
1415         // Dirty hack, for fixing disres and orires should be made mdmodules
1416         fr->fcdata->disres = disresdata;
1417         fr->fcdata->orires = oriresdata;
1418
1419         // Save a handle to device stream manager to use elsewhere in the code
1420         // TODO: Forcerec is not a correct place to store it.
1421         fr->deviceStreamManager = deviceStreamManager.get();
1422
1423         if (runScheduleWork.simulationWork.useGpuPmePpCommunication && !thisRankHasDuty(cr, DUTY_PME))
1424         {
1425             GMX_RELEASE_ASSERT(
1426                     deviceStreamManager != nullptr,
1427                     "GPU device stream manager should be valid in order to use PME-PP direct "
1428                     "communications.");
1429             GMX_RELEASE_ASSERT(
1430                     deviceStreamManager->streamIsValid(DeviceStreamType::PmePpTransfer),
1431                     "GPU PP-PME stream should be valid in order to use GPU PME-PP direct "
1432                     "communications.");
1433             fr->pmePpCommGpu = std::make_unique<gmx::PmePpCommGpu>(
1434                     cr->mpi_comm_mysim, cr->dd->pme_nodeid, deviceStreamManager->context(),
1435                     deviceStreamManager->stream(DeviceStreamType::PmePpTransfer));
1436         }
1437
1438         fr->nbv = Nbnxm::init_nb_verlet(mdlog, inputrec.get(), fr, cr, *hwinfo,
1439                                         runScheduleWork.simulationWork.useGpuNonbonded,
1440                                         deviceStreamManager.get(), &mtop, box, wcycle);
1441         // TODO: Move the logic below to a GPU bonded builder
1442         if (runScheduleWork.simulationWork.useGpuBonded)
1443         {
1444             GMX_RELEASE_ASSERT(deviceStreamManager != nullptr,
1445                                "GPU device stream manager should be valid in order to use GPU "
1446                                "version of bonded forces.");
1447             gpuBonded = std::make_unique<GpuBonded>(
1448                     mtop.ffparams, fr->ic->epsfac * fr->fudgeQQ, deviceStreamManager->context(),
1449                     deviceStreamManager->bondedStream(havePPDomainDecomposition(cr)), wcycle);
1450             fr->gpuBonded = gpuBonded.get();
1451         }
1452
1453         /* Initialize the mdAtoms structure.
1454          * mdAtoms is not filled with atom data,
1455          * as this can not be done now with domain decomposition.
1456          */
1457         mdAtoms = makeMDAtoms(fplog, mtop, *inputrec, thisRankHasPmeGpuTask);
1458         if (globalState && thisRankHasPmeGpuTask)
1459         {
1460             // The pinning of coordinates in the global state object works, because we only use
1461             // PME on GPU without DD or on a separate PME rank, and because the local state pointer
1462             // points to the global state object without DD.
1463             // FIXME: MD and EM separately set up the local state - this should happen in the same
1464             // function, which should also perform the pinning.
1465             changePinningPolicy(&globalState->x, pme_get_pinning_policy());
1466         }
1467
1468         /* Initialize the virtual site communication */
1469         vsite = makeVirtualSitesHandler(mtop, cr, fr->pbcType);
1470
1471         calc_shifts(box, fr->shift_vec);
1472
1473         /* With periodic molecules the charge groups should be whole at start up
1474          * and the virtual sites should not be far from their proper positions.
1475          */
1476         if (!inputrec->bContinuation && MASTER(cr)
1477             && !(inputrec->pbcType != PbcType::No && inputrec->bPeriodicMols))
1478         {
1479             /* Make molecules whole at start of run */
1480             if (fr->pbcType != PbcType::No)
1481             {
1482                 do_pbc_first_mtop(fplog, inputrec->pbcType, box, &mtop, globalState->x.rvec_array());
1483             }
1484             if (vsite)
1485             {
1486                 /* Correct initial vsite positions are required
1487                  * for the initial distribution in the domain decomposition
1488                  * and for the initial shell prediction.
1489                  */
1490                 constructVirtualSitesGlobal(mtop, globalState->x);
1491             }
1492         }
1493
1494         if (EEL_PME(fr->ic->eeltype) || EVDW_PME(fr->ic->vdwtype))
1495         {
1496             ewaldcoeff_q  = fr->ic->ewaldcoeff_q;
1497             ewaldcoeff_lj = fr->ic->ewaldcoeff_lj;
1498         }
1499     }
1500     else
1501     {
1502         /* This is a PME only node */
1503
1504         GMX_ASSERT(globalState == nullptr,
1505                    "We don't need the state on a PME only rank and expect it to be unitialized");
1506
1507         ewaldcoeff_q  = calc_ewaldcoeff_q(inputrec->rcoulomb, inputrec->ewald_rtol);
1508         ewaldcoeff_lj = calc_ewaldcoeff_lj(inputrec->rvdw, inputrec->ewald_rtol_lj);
1509     }
1510
1511     gmx_pme_t* sepPmeData = nullptr;
1512     // This reference hides the fact that PME data is owned by runner on PME-only ranks and by forcerec on other ranks
1513     GMX_ASSERT(thisRankHasDuty(cr, DUTY_PP) == (fr != nullptr),
1514                "Double-checking that only PME-only ranks have no forcerec");
1515     gmx_pme_t*& pmedata = fr ? fr->pmedata : sepPmeData;
1516
1517     // TODO should live in ewald module once its testing is improved
1518     //
1519     // Later, this program could contain kernels that might be later
1520     // re-used as auto-tuning progresses, or subsequent simulations
1521     // are invoked.
1522     PmeGpuProgramStorage pmeGpuProgram;
1523     if (thisRankHasPmeGpuTask)
1524     {
1525         GMX_RELEASE_ASSERT(
1526                 (deviceStreamManager != nullptr),
1527                 "GPU device stream manager should be initialized in order to use GPU for PME.");
1528         GMX_RELEASE_ASSERT((deviceInfo != nullptr),
1529                            "GPU device should be initialized in order to use GPU for PME.");
1530         pmeGpuProgram = buildPmeGpuProgram(deviceStreamManager->context());
1531     }
1532
1533     /* Initiate PME if necessary,
1534      * either on all nodes or on dedicated PME nodes only. */
1535     if (EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype))
1536     {
1537         if (mdAtoms && mdAtoms->mdatoms())
1538         {
1539             nChargePerturbed = mdAtoms->mdatoms()->nChargePerturbed;
1540             if (EVDW_PME(inputrec->vdwtype))
1541             {
1542                 nTypePerturbed = mdAtoms->mdatoms()->nTypePerturbed;
1543             }
1544         }
1545         if (cr->npmenodes > 0)
1546         {
1547             /* The PME only nodes need to know nChargePerturbed(FEP on Q) and nTypePerturbed(FEP on LJ)*/
1548             gmx_bcast(sizeof(nChargePerturbed), &nChargePerturbed, cr->mpi_comm_mysim);
1549             gmx_bcast(sizeof(nTypePerturbed), &nTypePerturbed, cr->mpi_comm_mysim);
1550         }
1551
1552         if (thisRankHasDuty(cr, DUTY_PME))
1553         {
1554             try
1555             {
1556                 // TODO: This should be in the builder.
1557                 GMX_RELEASE_ASSERT(!runScheduleWork.simulationWork.useGpuPme
1558                                            || (deviceStreamManager != nullptr),
1559                                    "Device stream manager should be valid in order to use GPU "
1560                                    "version of PME.");
1561                 GMX_RELEASE_ASSERT(
1562                         !runScheduleWork.simulationWork.useGpuPme
1563                                 || deviceStreamManager->streamIsValid(DeviceStreamType::Pme),
1564                         "GPU PME stream should be valid in order to use GPU version of PME.");
1565
1566                 const DeviceContext* deviceContext = runScheduleWork.simulationWork.useGpuPme
1567                                                              ? &deviceStreamManager->context()
1568                                                              : nullptr;
1569                 const DeviceStream* pmeStream =
1570                         runScheduleWork.simulationWork.useGpuPme
1571                                 ? &deviceStreamManager->stream(DeviceStreamType::Pme)
1572                                 : nullptr;
1573
1574                 pmedata = gmx_pme_init(cr, getNumPmeDomains(cr->dd), inputrec.get(),
1575                                        nChargePerturbed != 0, nTypePerturbed != 0,
1576                                        mdrunOptions.reproducible, ewaldcoeff_q, ewaldcoeff_lj,
1577                                        gmx_omp_nthreads_get(emntPME), pmeRunMode, nullptr,
1578                                        deviceContext, pmeStream, pmeGpuProgram.get(), mdlog);
1579             }
1580             GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1581         }
1582     }
1583
1584
1585     if (EI_DYNAMICS(inputrec->eI))
1586     {
1587         /* Turn on signal handling on all nodes */
1588         /*
1589          * (A user signal from the PME nodes (if any)
1590          * is communicated to the PP nodes.
1591          */
1592         signal_handler_install();
1593     }
1594
1595     pull_t* pull_work = nullptr;
1596     if (thisRankHasDuty(cr, DUTY_PP))
1597     {
1598         /* Assumes uniform use of the number of OpenMP threads */
1599         walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(emntDefault));
1600
1601         if (inputrec->bPull)
1602         {
1603             /* Initialize pull code */
1604             pull_work = init_pull(fplog, inputrec->pull, inputrec.get(), &mtop, cr, &atomSets,
1605                                   inputrec->fepvals->init_lambda);
1606             if (inputrec->pull->bXOutAverage || inputrec->pull->bFOutAverage)
1607             {
1608                 initPullHistory(pull_work, &observablesHistory);
1609             }
1610             if (EI_DYNAMICS(inputrec->eI) && MASTER(cr))
1611             {
1612                 init_pull_output_files(pull_work, filenames.size(), filenames.data(), oenv, startingBehavior);
1613             }
1614         }
1615
1616         std::unique_ptr<EnforcedRotation> enforcedRotation;
1617         if (inputrec->bRot)
1618         {
1619             /* Initialize enforced rotation code */
1620             enforcedRotation = init_rot(fplog, inputrec.get(), filenames.size(), filenames.data(),
1621                                         cr, &atomSets, globalState.get(), &mtop, oenv, mdrunOptions,
1622                                         startingBehavior);
1623         }
1624
1625         t_swap* swap = nullptr;
1626         if (inputrec->eSwapCoords != eswapNO)
1627         {
1628             /* Initialize ion swapping code */
1629             swap = init_swapcoords(fplog, inputrec.get(),
1630                                    opt2fn_master("-swap", filenames.size(), filenames.data(), cr),
1631                                    &mtop, globalState.get(), &observablesHistory, cr, &atomSets,
1632                                    oenv, mdrunOptions, startingBehavior);
1633         }
1634
1635         /* Let makeConstraints know whether we have essential dynamics constraints. */
1636         auto constr = makeConstraints(mtop, *inputrec, pull_work, doEssentialDynamics, fplog, cr,
1637                                       ms, &nrnb, wcycle, fr->bMolPBC);
1638
1639         /* Energy terms and groups */
1640         gmx_enerdata_t enerd(mtop.groups.groups[SimulationAtomGroupType::EnergyOutput].size(),
1641                              inputrec->fepvals->n_lambda);
1642
1643         /* Kinetic energy data */
1644         gmx_ekindata_t ekind;
1645         init_ekindata(fplog, &mtop, &(inputrec->opts), &ekind, inputrec->cos_accel);
1646
1647         /* Set up interactive MD (IMD) */
1648         auto imdSession =
1649                 makeImdSession(inputrec.get(), cr, wcycle, &enerd, ms, &mtop, mdlog,
1650                                MASTER(cr) ? globalState->x.rvec_array() : nullptr, filenames.size(),
1651                                filenames.data(), oenv, mdrunOptions.imdOptions, startingBehavior);
1652
1653         if (DOMAINDECOMP(cr))
1654         {
1655             GMX_RELEASE_ASSERT(fr, "fr was NULL while cr->duty was DUTY_PP");
1656             /* This call is not included in init_domain_decomposition mainly
1657              * because fr->cginfo_mb is set later.
1658              */
1659             dd_init_bondeds(fplog, cr->dd, mtop, vsite.get(), inputrec.get(),
1660                             domdecOptions.checkBondedInteractions, fr->cginfo_mb);
1661         }
1662
1663         std::unique_ptr<gmx::StatePropagatorDataGpu> stateGpu;
1664         if (gpusWereDetected
1665             && ((runScheduleWork.simulationWork.useGpuPme && thisRankHasDuty(cr, DUTY_PME))
1666                 || runScheduleWork.simulationWork.useGpuBufferOps))
1667         {
1668             GpuApiCallBehavior transferKind = (inputrec->eI == eiMD && !doRerun && !useModularSimulator)
1669                                                       ? GpuApiCallBehavior::Async
1670                                                       : GpuApiCallBehavior::Sync;
1671             GMX_RELEASE_ASSERT(deviceStreamManager != nullptr,
1672                                "GPU device stream manager should be initialized to use GPU.");
1673             stateGpu = std::make_unique<gmx::StatePropagatorDataGpu>(
1674                     *deviceStreamManager, transferKind, pme_gpu_get_block_size(fr->pmedata), wcycle);
1675             fr->stateGpu = stateGpu.get();
1676         }
1677
1678         GMX_ASSERT(stopHandlerBuilder_, "Runner must provide StopHandlerBuilder to simulator.");
1679         SimulatorBuilder simulatorBuilder;
1680
1681         simulatorBuilder.add(SimulatorStateData(globalState.get(), &observablesHistory, &enerd, &ekind));
1682         simulatorBuilder.add(std::move(membedHolder));
1683         simulatorBuilder.add(std::move(stopHandlerBuilder_));
1684         simulatorBuilder.add(SimulatorConfig(mdrunOptions, startingBehavior, &runScheduleWork));
1685
1686
1687         simulatorBuilder.add(SimulatorEnv(fplog, cr, ms, mdlog, oenv));
1688         simulatorBuilder.add(Profiling(&nrnb, walltime_accounting, wcycle));
1689         simulatorBuilder.add(ConstraintsParam(
1690                 constr.get(), enforcedRotation ? enforcedRotation->getLegacyEnfrot() : nullptr,
1691                 vsite.get()));
1692         // TODO: Separate `fr` to a separate add, and make the `build` handle the coupling sensibly.
1693         simulatorBuilder.add(LegacyInput(static_cast<int>(filenames.size()), filenames.data(),
1694                                          inputrec.get(), fr));
1695         simulatorBuilder.add(ReplicaExchangeParameters(replExParams));
1696         simulatorBuilder.add(InteractiveMD(imdSession.get()));
1697         simulatorBuilder.add(SimulatorModules(mdModules_->outputProvider(), mdModules_->notifier()));
1698         simulatorBuilder.add(CenterOfMassPulling(pull_work));
1699         // Todo move to an MDModule
1700         simulatorBuilder.add(IonSwapping(swap));
1701         simulatorBuilder.add(TopologyData(&mtop, mdAtoms.get()));
1702         simulatorBuilder.add(BoxDeformationHandle(deform.get()));
1703         simulatorBuilder.add(std::move(modularSimulatorCheckpointData));
1704
1705         // build and run simulator object based on user-input
1706         auto simulator = simulatorBuilder.build(useModularSimulator);
1707         simulator->run();
1708
1709         if (fr->pmePpCommGpu)
1710         {
1711             // destroy object since it is no longer required. (This needs to be done while the GPU context still exists.)
1712             fr->pmePpCommGpu.reset();
1713         }
1714
1715         if (inputrec->bPull)
1716         {
1717             finish_pull(pull_work);
1718         }
1719         finish_swapcoords(swap);
1720     }
1721     else
1722     {
1723         GMX_RELEASE_ASSERT(pmedata, "pmedata was NULL while cr->duty was not DUTY_PP");
1724         /* do PME only */
1725         walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(emntPME));
1726         gmx_pmeonly(pmedata, cr, &nrnb, wcycle, walltime_accounting, inputrec.get(), pmeRunMode,
1727                     deviceStreamManager.get());
1728     }
1729
1730     wallcycle_stop(wcycle, ewcRUN);
1731
1732     /* Finish up, write some stuff
1733      * if rerunMD, don't write last frame again
1734      */
1735     finish_run(fplog, mdlog, cr, inputrec.get(), &nrnb, wcycle, walltime_accounting,
1736                fr ? fr->nbv.get() : nullptr, pmedata, EI_DYNAMICS(inputrec->eI) && !isMultiSim(ms));
1737
1738     // clean up cycle counter
1739     wallcycle_destroy(wcycle);
1740
1741     deviceStreamManager.reset(nullptr);
1742     // Free PME data
1743     if (pmedata)
1744     {
1745         gmx_pme_destroy(pmedata);
1746         pmedata = nullptr;
1747     }
1748
1749     // FIXME: this is only here to manually unpin mdAtoms->chargeA_ and state->x,
1750     // before we destroy the GPU context(s)
1751     // Pinned buffers are associated with contexts in CUDA.
1752     // As soon as we destroy GPU contexts after mdrunner() exits, these lines should go.
1753     mdAtoms.reset(nullptr);
1754     globalState.reset(nullptr);
1755     mdModules_.reset(nullptr); // destruct force providers here as they might also use the GPU
1756     gpuBonded.reset(nullptr);
1757     /* Free pinned buffers in *fr */
1758     delete fr;
1759     fr = nullptr;
1760     // TODO convert to C++ so we can get rid of these frees
1761     sfree(disresdata);
1762     sfree(oriresdata);
1763
1764     if (!hwinfo->deviceInfoList.empty())
1765     {
1766         /* stop the GPU profiler (only CUDA) */
1767         stopGpuProfiler();
1768     }
1769
1770     /* With tMPI we need to wait for all ranks to finish deallocation before
1771      * destroying the CUDA context as some tMPI ranks may be sharing
1772      * GPU and context.
1773      *
1774      * This is not a concern in OpenCL where we use one context per rank.
1775      *
1776      * Note: it is safe to not call the barrier on the ranks which do not use GPU,
1777      * but it is easier and more futureproof to call it on the whole node.
1778      *
1779      * Note that this function needs to be called even if GPUs are not used
1780      * in this run because the PME ranks have no knowledge of whether GPUs
1781      * are used or not, but all ranks need to enter the barrier below.
1782      * \todo Remove this physical node barrier after making sure
1783      * that it's not needed anymore (with a shared GPU run).
1784      */
1785     if (GMX_THREAD_MPI)
1786     {
1787         physicalNodeComm.barrier();
1788     }
1789     releaseDevice(deviceInfo);
1790
1791     /* Does what it says */
1792     print_date_and_time(fplog, cr->nodeid, "Finished mdrun", gmx_gettime());
1793     walltime_accounting_destroy(walltime_accounting);
1794
1795     // Ensure log file content is written
1796     if (logFileHandle)
1797     {
1798         gmx_fio_flush(logFileHandle);
1799     }
1800
1801     /* Reset FPEs (important for unit tests) by disabling them. Assumes no
1802      * exceptions were enabled before function was called. */
1803     if (bEnableFPE)
1804     {
1805         gmx_fedisableexcept();
1806     }
1807
1808     auto rc = static_cast<int>(gmx_get_stop_condition());
1809
1810 #if GMX_THREAD_MPI
1811     /* we need to join all threads. The sub-threads join when they
1812        exit this function, but the master thread needs to be told to
1813        wait for that. */
1814     if (MASTER(cr))
1815     {
1816         tMPI_Finalize();
1817     }
1818 #endif
1819     return rc;
1820 } // namespace gmx
1821
1822 Mdrunner::~Mdrunner()
1823 {
1824     // Clean up of the Manager.
1825     // This will end up getting called on every thread-MPI rank, which is unnecessary,
1826     // but okay as long as threads synchronize some time before adding or accessing
1827     // a new set of restraints.
1828     if (restraintManager_)
1829     {
1830         restraintManager_->clear();
1831         GMX_ASSERT(restraintManager_->countRestraints() == 0,
1832                    "restraints added during runner life time should be cleared at runner "
1833                    "destruction.");
1834     }
1835 };
1836
1837 void Mdrunner::addPotential(std::shared_ptr<gmx::IRestraintPotential> puller, const std::string& name)
1838 {
1839     GMX_ASSERT(restraintManager_, "Mdrunner must have a restraint manager.");
1840     // Not sure if this should be logged through the md logger or something else,
1841     // but it is helpful to have some sort of INFO level message sent somewhere.
1842     //    std::cout << "Registering restraint named " << name << std::endl;
1843
1844     // When multiple restraints are used, it may be wasteful to register them separately.
1845     // Maybe instead register an entire Restraint Manager as a force provider.
1846     restraintManager_->addToSpec(std::move(puller), name);
1847 }
1848
1849 Mdrunner::Mdrunner(std::unique_ptr<MDModules> mdModules) : mdModules_(std::move(mdModules)) {}
1850
1851 Mdrunner::Mdrunner(Mdrunner&&) noexcept = default;
1852
1853 Mdrunner& Mdrunner::operator=(Mdrunner&& /*handle*/) noexcept = default;
1854
1855 class Mdrunner::BuilderImplementation
1856 {
1857 public:
1858     BuilderImplementation() = delete;
1859     BuilderImplementation(std::unique_ptr<MDModules> mdModules, compat::not_null<SimulationContext*> context);
1860     ~BuilderImplementation();
1861
1862     BuilderImplementation& setExtraMdrunOptions(const MdrunOptions& options,
1863                                                 real                forceWarningThreshold,
1864                                                 StartingBehavior    startingBehavior);
1865
1866     void addDomdec(const DomdecOptions& options);
1867
1868     void addInput(SimulationInputHandle inputHolder);
1869
1870     void addVerletList(int nstlist);
1871
1872     void addReplicaExchange(const ReplicaExchangeParameters& params);
1873
1874     void addNonBonded(const char* nbpu_opt);
1875
1876     void addPME(const char* pme_opt_, const char* pme_fft_opt_);
1877
1878     void addBondedTaskAssignment(const char* bonded_opt);
1879
1880     void addUpdateTaskAssignment(const char* update_opt);
1881
1882     void addHardwareOptions(const gmx_hw_opt_t& hardwareOptions);
1883
1884     void addFilenames(ArrayRef<const t_filenm> filenames);
1885
1886     void addOutputEnvironment(gmx_output_env_t* outputEnvironment);
1887
1888     void addLogFile(t_fileio* logFileHandle);
1889
1890     void addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder);
1891
1892     Mdrunner build();
1893
1894 private:
1895     // Default parameters copied from runner.h
1896     // \todo Clarify source(s) of default parameters.
1897
1898     const char* nbpu_opt_    = nullptr;
1899     const char* pme_opt_     = nullptr;
1900     const char* pme_fft_opt_ = nullptr;
1901     const char* bonded_opt_  = nullptr;
1902     const char* update_opt_  = nullptr;
1903
1904     MdrunOptions mdrunOptions_;
1905
1906     DomdecOptions domdecOptions_;
1907
1908     ReplicaExchangeParameters replicaExchangeParameters_;
1909
1910     //! Command-line override for the duration of a neighbor list with the Verlet scheme.
1911     int nstlist_ = 0;
1912
1913     //! World communicator, used for hardware detection and task assignment
1914     MPI_Comm worldCommunicator_ = MPI_COMM_NULL;
1915
1916     //! Multisim communicator handle.
1917     gmx_multisim_t* multiSimulation_;
1918
1919     //! mdrun communicator
1920     MPI_Comm communicator_ = MPI_COMM_NULL;
1921
1922     //! Print a warning if any force is larger than this (in kJ/mol nm).
1923     real forceWarningThreshold_ = -1;
1924
1925     //! Whether the simulation will start afresh, or restart with/without appending.
1926     StartingBehavior startingBehavior_ = StartingBehavior::NewSimulation;
1927
1928     //! The modules that comprise the functionality of mdrun.
1929     std::unique_ptr<MDModules> mdModules_;
1930
1931     //! \brief Parallelism information.
1932     gmx_hw_opt_t hardwareOptions_;
1933
1934     //! filename options for simulation.
1935     ArrayRef<const t_filenm> filenames_;
1936
1937     /*! \brief Handle to output environment.
1938      *
1939      * \todo gmx_output_env_t needs lifetime management.
1940      */
1941     gmx_output_env_t* outputEnvironment_ = nullptr;
1942
1943     /*! \brief Non-owning handle to MD log file.
1944      *
1945      * \todo Context should own output facilities for client.
1946      * \todo Improve log file handle management.
1947      * \internal
1948      * Code managing the FILE* relies on the ability to set it to
1949      * nullptr to check whether the filehandle is valid.
1950      */
1951     t_fileio* logFileHandle_ = nullptr;
1952
1953     /*!
1954      * \brief Builder for simulation stop signal handler.
1955      */
1956     std::unique_ptr<StopHandlerBuilder> stopHandlerBuilder_ = nullptr;
1957
1958     SimulationInputHandle inputHolder_;
1959 };
1960
1961 Mdrunner::BuilderImplementation::BuilderImplementation(std::unique_ptr<MDModules> mdModules,
1962                                                        compat::not_null<SimulationContext*> context) :
1963     mdModules_(std::move(mdModules))
1964 {
1965     worldCommunicator_ = context->worldCommunicator_;
1966     communicator_      = context->simulationCommunicator_;
1967     multiSimulation_   = context->multiSimulation_.get();
1968 }
1969
1970 Mdrunner::BuilderImplementation::~BuilderImplementation() = default;
1971
1972 Mdrunner::BuilderImplementation&
1973 Mdrunner::BuilderImplementation::setExtraMdrunOptions(const MdrunOptions&    options,
1974                                                       const real             forceWarningThreshold,
1975                                                       const StartingBehavior startingBehavior)
1976 {
1977     mdrunOptions_          = options;
1978     forceWarningThreshold_ = forceWarningThreshold;
1979     startingBehavior_      = startingBehavior;
1980     return *this;
1981 }
1982
1983 void Mdrunner::BuilderImplementation::addDomdec(const DomdecOptions& options)
1984 {
1985     domdecOptions_ = options;
1986 }
1987
1988 void Mdrunner::BuilderImplementation::addVerletList(int nstlist)
1989 {
1990     nstlist_ = nstlist;
1991 }
1992
1993 void Mdrunner::BuilderImplementation::addReplicaExchange(const ReplicaExchangeParameters& params)
1994 {
1995     replicaExchangeParameters_ = params;
1996 }
1997
1998 Mdrunner Mdrunner::BuilderImplementation::build()
1999 {
2000     auto newRunner = Mdrunner(std::move(mdModules_));
2001
2002     newRunner.mdrunOptions     = mdrunOptions_;
2003     newRunner.pforce           = forceWarningThreshold_;
2004     newRunner.startingBehavior = startingBehavior_;
2005     newRunner.domdecOptions    = domdecOptions_;
2006
2007     // \todo determine an invariant to check or confirm that all gmx_hw_opt_t objects are valid
2008     newRunner.hw_opt = hardwareOptions_;
2009
2010     // No invariant to check. This parameter exists to optionally override other behavior.
2011     newRunner.nstlist_cmdline = nstlist_;
2012
2013     newRunner.replExParams = replicaExchangeParameters_;
2014
2015     newRunner.filenames = filenames_;
2016
2017     newRunner.worldCommunicator = worldCommunicator_;
2018
2019     newRunner.communicator = communicator_;
2020
2021     // nullptr is a valid value for the multisim handle
2022     newRunner.ms = multiSimulation_;
2023
2024     if (inputHolder_)
2025     {
2026         newRunner.inputHolder_ = std::move(inputHolder_);
2027     }
2028     else
2029     {
2030         GMX_THROW(gmx::APIError("MdrunnerBuilder::addInput() is required before build()."));
2031     }
2032
2033     // \todo Clarify ownership and lifetime management for gmx_output_env_t
2034     // \todo Update sanity checking when output environment has clearly specified invariants.
2035     // Initialization and default values for oenv are not well specified in the current version.
2036     if (outputEnvironment_)
2037     {
2038         newRunner.oenv = outputEnvironment_;
2039     }
2040     else
2041     {
2042         GMX_THROW(gmx::APIError(
2043                 "MdrunnerBuilder::addOutputEnvironment() is required before build()"));
2044     }
2045
2046     newRunner.logFileHandle = logFileHandle_;
2047
2048     if (nbpu_opt_)
2049     {
2050         newRunner.nbpu_opt = nbpu_opt_;
2051     }
2052     else
2053     {
2054         GMX_THROW(gmx::APIError("MdrunnerBuilder::addNonBonded() is required before build()"));
2055     }
2056
2057     if (pme_opt_ && pme_fft_opt_)
2058     {
2059         newRunner.pme_opt     = pme_opt_;
2060         newRunner.pme_fft_opt = pme_fft_opt_;
2061     }
2062     else
2063     {
2064         GMX_THROW(gmx::APIError("MdrunnerBuilder::addElectrostatics() is required before build()"));
2065     }
2066
2067     if (bonded_opt_)
2068     {
2069         newRunner.bonded_opt = bonded_opt_;
2070     }
2071     else
2072     {
2073         GMX_THROW(gmx::APIError(
2074                 "MdrunnerBuilder::addBondedTaskAssignment() is required before build()"));
2075     }
2076
2077     if (update_opt_)
2078     {
2079         newRunner.update_opt = update_opt_;
2080     }
2081     else
2082     {
2083         GMX_THROW(gmx::APIError(
2084                 "MdrunnerBuilder::addUpdateTaskAssignment() is required before build()  "));
2085     }
2086
2087
2088     newRunner.restraintManager_ = std::make_unique<gmx::RestraintManager>();
2089
2090     if (stopHandlerBuilder_)
2091     {
2092         newRunner.stopHandlerBuilder_ = std::move(stopHandlerBuilder_);
2093     }
2094     else
2095     {
2096         newRunner.stopHandlerBuilder_ = std::make_unique<StopHandlerBuilder>();
2097     }
2098
2099     return newRunner;
2100 }
2101
2102 void Mdrunner::BuilderImplementation::addNonBonded(const char* nbpu_opt)
2103 {
2104     nbpu_opt_ = nbpu_opt;
2105 }
2106
2107 void Mdrunner::BuilderImplementation::addPME(const char* pme_opt, const char* pme_fft_opt)
2108 {
2109     pme_opt_     = pme_opt;
2110     pme_fft_opt_ = pme_fft_opt;
2111 }
2112
2113 void Mdrunner::BuilderImplementation::addBondedTaskAssignment(const char* bonded_opt)
2114 {
2115     bonded_opt_ = bonded_opt;
2116 }
2117
2118 void Mdrunner::BuilderImplementation::addUpdateTaskAssignment(const char* update_opt)
2119 {
2120     update_opt_ = update_opt;
2121 }
2122
2123 void Mdrunner::BuilderImplementation::addHardwareOptions(const gmx_hw_opt_t& hardwareOptions)
2124 {
2125     hardwareOptions_ = hardwareOptions;
2126 }
2127
2128 void Mdrunner::BuilderImplementation::addFilenames(ArrayRef<const t_filenm> filenames)
2129 {
2130     filenames_ = filenames;
2131 }
2132
2133 void Mdrunner::BuilderImplementation::addOutputEnvironment(gmx_output_env_t* outputEnvironment)
2134 {
2135     outputEnvironment_ = outputEnvironment;
2136 }
2137
2138 void Mdrunner::BuilderImplementation::addLogFile(t_fileio* logFileHandle)
2139 {
2140     logFileHandle_ = logFileHandle;
2141 }
2142
2143 void Mdrunner::BuilderImplementation::addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder)
2144 {
2145     stopHandlerBuilder_ = std::move(builder);
2146 }
2147
2148 void Mdrunner::BuilderImplementation::addInput(SimulationInputHandle inputHolder)
2149 {
2150     inputHolder_ = std::move(inputHolder);
2151 }
2152
2153 MdrunnerBuilder::MdrunnerBuilder(std::unique_ptr<MDModules>           mdModules,
2154                                  compat::not_null<SimulationContext*> context) :
2155     impl_{ std::make_unique<Mdrunner::BuilderImplementation>(std::move(mdModules), context) }
2156 {
2157 }
2158
2159 MdrunnerBuilder::~MdrunnerBuilder() = default;
2160
2161 MdrunnerBuilder& MdrunnerBuilder::addSimulationMethod(const MdrunOptions&    options,
2162                                                       real                   forceWarningThreshold,
2163                                                       const StartingBehavior startingBehavior)
2164 {
2165     impl_->setExtraMdrunOptions(options, forceWarningThreshold, startingBehavior);
2166     return *this;
2167 }
2168
2169 MdrunnerBuilder& MdrunnerBuilder::addDomainDecomposition(const DomdecOptions& options)
2170 {
2171     impl_->addDomdec(options);
2172     return *this;
2173 }
2174
2175 MdrunnerBuilder& MdrunnerBuilder::addNeighborList(int nstlist)
2176 {
2177     impl_->addVerletList(nstlist);
2178     return *this;
2179 }
2180
2181 MdrunnerBuilder& MdrunnerBuilder::addReplicaExchange(const ReplicaExchangeParameters& params)
2182 {
2183     impl_->addReplicaExchange(params);
2184     return *this;
2185 }
2186
2187 MdrunnerBuilder& MdrunnerBuilder::addNonBonded(const char* nbpu_opt)
2188 {
2189     impl_->addNonBonded(nbpu_opt);
2190     return *this;
2191 }
2192
2193 MdrunnerBuilder& MdrunnerBuilder::addElectrostatics(const char* pme_opt, const char* pme_fft_opt)
2194 {
2195     // The builder method may become more general in the future, but in this version,
2196     // parameters for PME electrostatics are both required and the only parameters
2197     // available.
2198     if (pme_opt && pme_fft_opt)
2199     {
2200         impl_->addPME(pme_opt, pme_fft_opt);
2201     }
2202     else
2203     {
2204         GMX_THROW(
2205                 gmx::InvalidInputError("addElectrostatics() arguments must be non-null pointers."));
2206     }
2207     return *this;
2208 }
2209
2210 MdrunnerBuilder& MdrunnerBuilder::addBondedTaskAssignment(const char* bonded_opt)
2211 {
2212     impl_->addBondedTaskAssignment(bonded_opt);
2213     return *this;
2214 }
2215
2216 MdrunnerBuilder& MdrunnerBuilder::addUpdateTaskAssignment(const char* update_opt)
2217 {
2218     impl_->addUpdateTaskAssignment(update_opt);
2219     return *this;
2220 }
2221
2222 Mdrunner MdrunnerBuilder::build()
2223 {
2224     return impl_->build();
2225 }
2226
2227 MdrunnerBuilder& MdrunnerBuilder::addHardwareOptions(const gmx_hw_opt_t& hardwareOptions)
2228 {
2229     impl_->addHardwareOptions(hardwareOptions);
2230     return *this;
2231 }
2232
2233 MdrunnerBuilder& MdrunnerBuilder::addFilenames(ArrayRef<const t_filenm> filenames)
2234 {
2235     impl_->addFilenames(filenames);
2236     return *this;
2237 }
2238
2239 MdrunnerBuilder& MdrunnerBuilder::addOutputEnvironment(gmx_output_env_t* outputEnvironment)
2240 {
2241     impl_->addOutputEnvironment(outputEnvironment);
2242     return *this;
2243 }
2244
2245 MdrunnerBuilder& MdrunnerBuilder::addLogFile(t_fileio* logFileHandle)
2246 {
2247     impl_->addLogFile(logFileHandle);
2248     return *this;
2249 }
2250
2251 MdrunnerBuilder& MdrunnerBuilder::addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder)
2252 {
2253     impl_->addStopHandlerBuilder(std::move(builder));
2254     return *this;
2255 }
2256
2257 MdrunnerBuilder& MdrunnerBuilder::addInput(SimulationInputHandle input)
2258 {
2259     impl_->addInput(std::move(input));
2260     return *this;
2261 }
2262
2263 MdrunnerBuilder::MdrunnerBuilder(MdrunnerBuilder&&) noexcept = default;
2264
2265 MdrunnerBuilder& MdrunnerBuilder::operator=(MdrunnerBuilder&&) noexcept = default;
2266
2267 } // namespace gmx