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) 2011-2019,2020,2021, by the GROMACS development team, led by
   7  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   8  * and including many others, as listed in the AUTHORS file in the
   9  * top-level source directory and at http://www.gromacs.org.
  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_stream_manager.h"
  77 #include "gromacs/hardware/cpuinfo.h"
  78 #include "gromacs/hardware/detecthardware.h"
  79 #include "gromacs/hardware/device_management.h"
  80 #include "gromacs/hardware/hardwaretopology.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/gpuforcereduction.h"
  99 #include "gromacs/mdlib/makeconstraints.h"
 100 #include "gromacs/mdlib/md_support.h"
 101 #include "gromacs/mdlib/mdatoms.h"
 102 #include "gromacs/mdlib/sighandler.h"
 103 #include "gromacs/mdlib/stophandler.h"
 104 #include "gromacs/mdlib/tgroup.h"
 105 #include "gromacs/mdlib/updategroups.h"
 106 #include "gromacs/mdlib/vsite.h"
 107 #include "gromacs/mdrun/mdmodules.h"
 108 #include "gromacs/mdrun/simulationcontext.h"
 109 #include "gromacs/mdrun/simulationinput.h"
 110 #include "gromacs/mdrun/simulationinputhandle.h"
 111 #include "gromacs/mdrunutility/handlerestart.h"
 112 #include "gromacs/mdrunutility/logging.h"
 113 #include "gromacs/mdrunutility/multisim.h"
 114 #include "gromacs/mdrunutility/printtime.h"
 115 #include "gromacs/mdrunutility/threadaffinity.h"
 116 #include "gromacs/mdtypes/checkpointdata.h"
 117 #include "gromacs/mdtypes/commrec.h"
 118 #include "gromacs/mdtypes/enerdata.h"
 119 #include "gromacs/mdtypes/fcdata.h"
 120 #include "gromacs/mdtypes/forcerec.h"
 121 #include "gromacs/mdtypes/group.h"
 122 #include "gromacs/mdtypes/inputrec.h"
 123 #include "gromacs/mdtypes/interaction_const.h"
 124 #include "gromacs/mdtypes/md_enums.h"
 125 #include "gromacs/mdtypes/mdatom.h"
 126 #include "gromacs/mdtypes/mdrunoptions.h"
 127 #include "gromacs/mdtypes/observableshistory.h"
 128 #include "gromacs/mdtypes/simulation_workload.h"
 129 #include "gromacs/mdtypes/state.h"
 130 #include "gromacs/mdtypes/state_propagator_data_gpu.h"
 131 #include "gromacs/modularsimulator/modularsimulator.h"
 132 #include "gromacs/nbnxm/gpu_data_mgmt.h"
 133 #include "gromacs/nbnxm/nbnxm.h"
 134 #include "gromacs/nbnxm/pairlist_tuning.h"
 135 #include "gromacs/pbcutil/pbc.h"
 136 #include "gromacs/pulling/output.h"
 137 #include "gromacs/pulling/pull.h"
 138 #include "gromacs/pulling/pull_rotation.h"
 139 #include "gromacs/restraint/manager.h"
 140 #include "gromacs/restraint/restraintmdmodule.h"
 141 #include "gromacs/restraint/restraintpotential.h"
 142 #include "gromacs/swap/swapcoords.h"
 143 #include "gromacs/taskassignment/decidegpuusage.h"
 144 #include "gromacs/taskassignment/decidesimulationworkload.h"
 145 #include "gromacs/taskassignment/resourcedivision.h"
 146 #include "gromacs/taskassignment/taskassignment.h"
 147 #include "gromacs/taskassignment/usergpuids.h"
 148 #include "gromacs/timing/gpu_timing.h"
 149 #include "gromacs/timing/wallcycle.h"
 150 #include "gromacs/timing/wallcyclereporting.h"
 151 #include "gromacs/topology/mtop_util.h"
 152 #include "gromacs/trajectory/trajectoryframe.h"
 153 #include "gromacs/utility/basenetwork.h"
 154 #include "gromacs/utility/cstringutil.h"
 155 #include "gromacs/utility/exceptions.h"
 156 #include "gromacs/utility/fatalerror.h"
 157 #include "gromacs/utility/filestream.h"
 158 #include "gromacs/utility/gmxassert.h"
 159 #include "gromacs/utility/gmxmpi.h"
 160 #include "gromacs/utility/keyvaluetree.h"
 161 #include "gromacs/utility/logger.h"
 162 #include "gromacs/utility/loggerbuilder.h"
 163 #include "gromacs/utility/mdmodulenotification.h"
 164 #include "gromacs/utility/physicalnodecommunicator.h"
 165 #include "gromacs/utility/pleasecite.h"
 166 #include "gromacs/utility/programcontext.h"
 167 #include "gromacs/utility/smalloc.h"
 168 #include "gromacs/utility/stringutil.h"
 169
 170 #include "isimulator.h"
 171 #include "membedholder.h"
 172 #include "replicaexchange.h"
 173 #include "simulatorbuilder.h"
 174
 175 namespace gmx
 176 {
 177
 178
 179 /*! \brief Manage any development feature flag variables encountered
 180  *
 181  * The use of dev features indicated by environment variables is
 182  * logged in order to ensure that runs with such features enabled can
 183  * be identified from their log and standard output. Any cross
 184  * dependencies are also checked, and if unsatisfied, a fatal error
 185  * issued.
 186  *
 187  * Note that some development features overrides are applied already here:
 188  * the GPU communication flags are set to false in non-tMPI and non-CUDA builds.
 189  *
 190  * \param[in]  mdlog                Logger object.
 191  * \param[in]  useGpuForNonbonded   True if the nonbonded task is offloaded in this run.
 192  * \param[in]  pmeRunMode           The PME run mode for this run
 193  * \returns                         The object populated with development feature flags.
 194  */
 195 static DevelopmentFeatureFlags manageDevelopmentFeatures(const gmx::MDLogger& mdlog,
 196                                                          const bool           useGpuForNonbonded,
 197                                                          const PmeRunMode     pmeRunMode)
 198 {
 199     DevelopmentFeatureFlags devFlags;
 200
 201     // Some builds of GCC 5 give false positive warnings that these
 202     // getenv results are ignored when clearly they are used.
 203 #pragma GCC diagnostic push
 204 #pragma GCC diagnostic ignored "-Wunused-result"
 205
 206     devFlags.enableGpuBufferOps =
 207             GMX_GPU_CUDA && useGpuForNonbonded && (getenv("GMX_USE_GPU_BUFFER_OPS") != nullptr);
 208     devFlags.enableGpuHaloExchange = GMX_GPU_CUDA && GMX_THREAD_MPI && getenv("GMX_GPU_DD_COMMS") != nullptr;
 209     devFlags.forceGpuUpdateDefault = (getenv("GMX_FORCE_UPDATE_DEFAULT_GPU") != nullptr) || GMX_FAHCORE;
 210     devFlags.enableGpuPmePPComm =
 211             GMX_GPU_CUDA && GMX_THREAD_MPI && getenv("GMX_GPU_PME_PP_COMMS") != nullptr;
 212
 213 #pragma GCC diagnostic pop
 214
 215     if (devFlags.enableGpuBufferOps)
 216     {
 217         GMX_LOG(mdlog.warning)
 218                 .asParagraph()
 219                 .appendTextFormatted(
 220                         "This run uses the 'GPU buffer ops' feature, enabled by the "
 221                         "GMX_USE_GPU_BUFFER_OPS environment variable.");
 222     }
 223
 224     if (devFlags.forceGpuUpdateDefault)
 225     {
 226         GMX_LOG(mdlog.warning)
 227                 .asParagraph()
 228                 .appendTextFormatted(
 229                         "This run will default to '-update gpu' as requested by the "
 230                         "GMX_FORCE_UPDATE_DEFAULT_GPU environment variable. GPU update with domain "
 231                         "decomposition lacks substantial testing and should be used with caution.");
 232     }
 233
 234     if (devFlags.enableGpuHaloExchange)
 235     {
 236         if (useGpuForNonbonded)
 237         {
 238             if (!devFlags.enableGpuBufferOps)
 239             {
 240                 GMX_LOG(mdlog.warning)
 241                         .asParagraph()
 242                         .appendTextFormatted(
 243                                 "Enabling GPU buffer operations required by GMX_GPU_DD_COMMS "
 244                                 "(equivalent with GMX_USE_GPU_BUFFER_OPS=1).");
 245                 devFlags.enableGpuBufferOps = true;
 246             }
 247             GMX_LOG(mdlog.warning)
 248                     .asParagraph()
 249                     .appendTextFormatted(
 250                             "This run has requested the 'GPU halo exchange' feature, enabled by "
 251                             "the "
 252                             "GMX_GPU_DD_COMMS environment variable.");
 253         }
 254         else
 255         {
 256             GMX_LOG(mdlog.warning)
 257                     .asParagraph()
 258                     .appendTextFormatted(
 259                             "GMX_GPU_DD_COMMS environment variable detected, but the 'GPU "
 260                             "halo exchange' feature will not be enabled as nonbonded interactions "
 261                             "are not offloaded.");
 262             devFlags.enableGpuHaloExchange = false;
 263         }
 264     }
 265
 266     if (devFlags.enableGpuPmePPComm)
 267     {
 268         if (pmeRunMode == PmeRunMode::GPU)
 269         {
 270             if (!devFlags.enableGpuBufferOps)
 271             {
 272                 GMX_LOG(mdlog.warning)
 273                         .asParagraph()
 274                         .appendTextFormatted(
 275                                 "Enabling GPU buffer operations required by GMX_GPU_PME_PP_COMMS "
 276                                 "(equivalent with GMX_USE_GPU_BUFFER_OPS=1).");
 277                 devFlags.enableGpuBufferOps = true;
 278             }
 279             GMX_LOG(mdlog.warning)
 280                     .asParagraph()
 281                     .appendTextFormatted(
 282                             "This run uses the 'GPU PME-PP communications' feature, enabled "
 283                             "by the GMX_GPU_PME_PP_COMMS environment variable.");
 284         }
 285         else
 286         {
 287             std::string clarification;
 288             if (pmeRunMode == PmeRunMode::Mixed)
 289             {
 290                 clarification =
 291                         "PME FFT and gather are not offloaded to the GPU (PME is running in mixed "
 292                         "mode).";
 293             }
 294             else
 295             {
 296                 clarification = "PME is not offloaded to the GPU.";
 297             }
 298             GMX_LOG(mdlog.warning)
 299                     .asParagraph()
 300                     .appendText(
 301                             "GMX_GPU_PME_PP_COMMS environment variable detected, but the "
 302                             "'GPU PME-PP communications' feature was not enabled as "
 303                             + clarification);
 304             devFlags.enableGpuPmePPComm = false;
 305         }
 306     }
 307
 308     return devFlags;
 309 }
 310
 311 /*! \brief Barrier for safe simultaneous thread access to mdrunner data
 312  *
 313  * Used to ensure that the master thread does not modify mdrunner during copy
 314  * on the spawned threads. */
 315 static void threadMpiMdrunnerAccessBarrier()
 316 {
 317 #if GMX_THREAD_MPI
 318     MPI_Barrier(MPI_COMM_WORLD);
 319 #endif
 320 }
 321
 322 Mdrunner Mdrunner::cloneOnSpawnedThread() const
 323 {
 324     auto newRunner = Mdrunner(std::make_unique<MDModules>());
 325
 326     // All runners in the same process share a restraint manager resource because it is
 327     // part of the interface to the client code, which is associated only with the
 328     // original thread. Handles to the same resources can be obtained by copy.
 329     {
 330         newRunner.restraintManager_ = std::make_unique<RestraintManager>(*restraintManager_);
 331     }
 332
 333     // Copy members of master runner.
 334     // \todo Replace with builder when Simulation context and/or runner phases are better defined.
 335     // Ref https://gitlab.com/gromacs/gromacs/-/issues/2587 and https://gitlab.com/gromacs/gromacs/-/issues/2375
 336     newRunner.hw_opt    = hw_opt;
 337     newRunner.filenames = filenames;
 338
 339     newRunner.hwinfo_         = hwinfo_;
 340     newRunner.oenv            = oenv;
 341     newRunner.mdrunOptions    = mdrunOptions;
 342     newRunner.domdecOptions   = domdecOptions;
 343     newRunner.nbpu_opt        = nbpu_opt;
 344     newRunner.pme_opt         = pme_opt;
 345     newRunner.pme_fft_opt     = pme_fft_opt;
 346     newRunner.bonded_opt      = bonded_opt;
 347     newRunner.update_opt      = update_opt;
 348     newRunner.nstlist_cmdline = nstlist_cmdline;
 349     newRunner.replExParams    = replExParams;
 350     newRunner.pforce          = pforce;
 351     // Give the spawned thread the newly created valid communicator
 352     // for the simulation.
 353     newRunner.libraryWorldCommunicator = MPI_COMM_WORLD;
 354     newRunner.simulationCommunicator   = MPI_COMM_WORLD;
 355     newRunner.ms                       = ms;
 356     newRunner.startingBehavior         = startingBehavior;
 357     newRunner.stopHandlerBuilder_      = std::make_unique<StopHandlerBuilder>(*stopHandlerBuilder_);
 358     newRunner.inputHolder_             = inputHolder_;
 359
 360     threadMpiMdrunnerAccessBarrier();
 361
 362     return newRunner;
 363 }
 364
 365 /*! \brief The callback used for running on spawned threads.
 366  *
 367  * Obtains the pointer to the master mdrunner object from the one
 368  * argument permitted to the thread-launch API call, copies it to make
 369  * a new runner for this thread, reinitializes necessary data, and
 370  * proceeds to the simulation. */
 371 static void mdrunner_start_fn(const void* arg)
 372 {
 373     try
 374     {
 375         auto masterMdrunner = reinterpret_cast<const gmx::Mdrunner*>(arg);
 376         /* copy the arg list to make sure that it's thread-local. This
 377            doesn't copy pointed-to items, of course; fnm, cr and fplog
 378            are reset in the call below, all others should be const. */
 379         gmx::Mdrunner mdrunner = masterMdrunner->cloneOnSpawnedThread();
 380         mdrunner.mdrunner();
 381     }
 382     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
 383 }
 384
 385
 386 void Mdrunner::spawnThreads(int numThreadsToLaunch)
 387 {
 388 #if GMX_THREAD_MPI
 389     /* now spawn new threads that start mdrunner_start_fn(), while
 390        the main thread returns. Thread affinity is handled later. */
 391     if (tMPI_Init_fn(TRUE, numThreadsToLaunch, TMPI_AFFINITY_NONE, mdrunner_start_fn, static_cast<const void*>(this))
 392         != TMPI_SUCCESS)
 393     {
 394         GMX_THROW(gmx::InternalError("Failed to spawn thread-MPI threads"));
 395     }
 396
 397     // Give the master thread the newly created valid communicator for
 398     // the simulation.
 399     libraryWorldCommunicator = MPI_COMM_WORLD;
 400     simulationCommunicator   = MPI_COMM_WORLD;
 401     threadMpiMdrunnerAccessBarrier();
 402 #else
 403     GMX_UNUSED_VALUE(numThreadsToLaunch);
 404     GMX_UNUSED_VALUE(mdrunner_start_fn);
 405 #endif
 406 }
 407
 408 } // namespace gmx
 409
 410 /*! \brief Initialize variables for Verlet scheme simulation */
 411 static void prepare_verlet_scheme(FILE*               fplog,
 412                                   t_commrec*          cr,
 413                                   t_inputrec*         ir,
 414                                   int                 nstlist_cmdline,
 415                                   const gmx_mtop_t*   mtop,
 416                                   const matrix        box,
 417                                   bool                makeGpuPairList,
 418                                   const gmx::CpuInfo& cpuinfo)
 419 {
 420     // We checked the cut-offs in grompp, but double-check here.
 421     // We have PME+LJcutoff kernels for rcoulomb>rvdw.
 422     if (EEL_PME_EWALD(ir->coulombtype) && ir->vdwtype == eelCUT)
 423     {
 424         GMX_RELEASE_ASSERT(ir->rcoulomb >= ir->rvdw,
 425                            "With Verlet lists and PME we should have rcoulomb>=rvdw");
 426     }
 427     else
 428     {
 429         GMX_RELEASE_ASSERT(ir->rcoulomb == ir->rvdw,
 430                            "With Verlet lists and no PME rcoulomb and rvdw should be identical");
 431     }
 432     /* For NVE simulations, we will retain the initial list buffer */
 433     if (EI_DYNAMICS(ir->eI) && ir->verletbuf_tol > 0 && !(EI_MD(ir->eI) && ir->etc == etcNO))
 434     {
 435         /* Update the Verlet buffer size for the current run setup */
 436
 437         /* Here we assume SIMD-enabled kernels are being used. But as currently
 438          * calc_verlet_buffer_size gives the same results for 4x8 and 4x4
 439          * and 4x2 gives a larger buffer than 4x4, this is ok.
 440          */
 441         ListSetupType listType =
 442                 (makeGpuPairList ? ListSetupType::Gpu : ListSetupType::CpuSimdWhenSupported);
 443         VerletbufListSetup listSetup = verletbufGetSafeListSetup(listType);
 444
 445         const real rlist_new =
 446                 calcVerletBufferSize(*mtop, det(box), *ir, ir->nstlist, ir->nstlist - 1, -1, listSetup);
 447
 448         if (rlist_new != ir->rlist)
 449         {
 450             if (fplog != nullptr)
 451             {
 452                 fprintf(fplog,
 453                         "\nChanging rlist from %g to %g for non-bonded %dx%d atom kernels\n\n",
 454                         ir->rlist,
 455                         rlist_new,
 456                         listSetup.cluster_size_i,
 457                         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,
 466                   "Can not set nstlist without %s",
 467                   !EI_DYNAMICS(ir->eI) ? "dynamics" : "verlet-buffer-tolerance");
 468     }
 469
 470     if (EI_DYNAMICS(ir->eI))
 471     {
 472         /* Set or try nstlist values */
 473         increaseNstlist(fplog, cr, ir, nstlist_cmdline, mtop, box, makeGpuPairList, cpuinfo);
 474     }
 475 }
 476
 477 /*! \brief Override the nslist value in inputrec
 478  *
 479  * with value passed on the command line (if any)
 480  */
 481 static void override_nsteps_cmdline(const gmx::MDLogger& mdlog, int64_t nsteps_cmdline, t_inputrec* ir)
 482 {
 483     assert(ir);
 484
 485     /* override with anything else than the default -2 */
 486     if (nsteps_cmdline > -2)
 487     {
 488         char sbuf_steps[STEPSTRSIZE];
 489         char sbuf_msg[STRLEN];
 490
 491         ir->nsteps = nsteps_cmdline;
 492         if (EI_DYNAMICS(ir->eI) && nsteps_cmdline != -1)
 493         {
 494             sprintf(sbuf_msg,
 495                     "Overriding nsteps with value passed on the command line: %s steps, %.3g ps",
 496                     gmx_step_str(nsteps_cmdline, sbuf_steps),
 497                     fabs(nsteps_cmdline * ir->delta_t));
 498         }
 499         else
 500         {
 501             sprintf(sbuf_msg,
 502                     "Overriding nsteps with value passed on the command line: %s steps",
 503                     gmx_step_str(nsteps_cmdline, sbuf_steps));
 504         }
 505
 506         GMX_LOG(mdlog.warning).asParagraph().appendText(sbuf_msg);
 507     }
 508     else if (nsteps_cmdline < -2)
 509     {
 510         gmx_fatal(FARGS, "Invalid nsteps value passed on the command line: %" PRId64, nsteps_cmdline);
 511     }
 512     /* Do nothing if nsteps_cmdline == -2 */
 513 }
 514
 515 namespace gmx
 516 {
 517
 518 /*! \brief Return whether GPU acceleration of nonbondeds is supported with the given settings.
 519  *
 520  * If not, and if a warning may be issued, logs a warning about
 521  * falling back to CPU code. With thread-MPI, only the first
 522  * call to this function should have \c issueWarning true. */
 523 static bool gpuAccelerationOfNonbondedIsUseful(const MDLogger& mdlog, const t_inputrec& ir, bool issueWarning)
 524 {
 525     bool        gpuIsUseful = true;
 526     std::string warning;
 527
 528     if (ir.opts.ngener - ir.nwall > 1)
 529     {
 530         /* The GPU code does not support more than one energy group.
 531          * If the user requested GPUs explicitly, a fatal error is given later.
 532          */
 533         gpuIsUseful = false;
 534         warning =
 535                 "Multiple energy groups is not implemented for GPUs, falling back to the CPU. "
 536                 "For better performance, run on the GPU without energy groups and then do "
 537                 "gmx mdrun -rerun option on the trajectory with an energy group .tpr file.";
 538     }
 539
 540     if (EI_TPI(ir.eI))
 541     {
 542         gpuIsUseful = false;
 543         warning     = "TPI is not implemented for GPUs.";
 544     }
 545
 546     if (!gpuIsUseful && issueWarning)
 547     {
 548         GMX_LOG(mdlog.warning).asParagraph().appendText(warning);
 549     }
 550
 551     return gpuIsUseful;
 552 }
 553
 554 //! Initializes the logger for mdrun.
 555 static gmx::LoggerOwner buildLogger(FILE* fplog, const bool isSimulationMasterRank)
 556 {
 557     gmx::LoggerBuilder builder;
 558     if (fplog != nullptr)
 559     {
 560         builder.addTargetFile(gmx::MDLogger::LogLevel::Info, fplog);
 561     }
 562     if (isSimulationMasterRank)
 563     {
 564         builder.addTargetStream(gmx::MDLogger::LogLevel::Warning, &gmx::TextOutputFile::standardError());
 565     }
 566     return builder.build();
 567 }
 568
 569 //! Make a TaskTarget from an mdrun argument string.
 570 static TaskTarget findTaskTarget(const char* optionString)
 571 {
 572     TaskTarget returnValue = TaskTarget::Auto;
 573
 574     if (strncmp(optionString, "auto", 3) == 0)
 575     {
 576         returnValue = TaskTarget::Auto;
 577     }
 578     else if (strncmp(optionString, "cpu", 3) == 0)
 579     {
 580         returnValue = TaskTarget::Cpu;
 581     }
 582     else if (strncmp(optionString, "gpu", 3) == 0)
 583     {
 584         returnValue = TaskTarget::Gpu;
 585     }
 586     else
 587     {
 588         GMX_ASSERT(false, "Option string should have been checked for sanity already");
 589     }
 590
 591     return returnValue;
 592 }
 593
 594 //! Finish run, aggregate data to print performance info.
 595 static void finish_run(FILE*                     fplog,
 596                        const gmx::MDLogger&      mdlog,
 597                        const t_commrec*          cr,
 598                        const t_inputrec*         inputrec,
 599                        t_nrnb                    nrnb[],
 600                        gmx_wallcycle_t           wcycle,
 601                        gmx_walltime_accounting_t walltime_accounting,
 602                        nonbonded_verlet_t*       nbv,
 603                        const gmx_pme_t*          pme,
 604                        gmx_bool                  bWriteStat)
 605 {
 606     double delta_t = 0;
 607     double nbfs = 0, mflop = 0;
 608     double elapsed_time, elapsed_time_over_all_ranks, elapsed_time_over_all_threads,
 609             elapsed_time_over_all_threads_over_all_ranks;
 610     /* Control whether it is valid to print a report. Only the
 611        simulation master may print, but it should not do so if the run
 612        terminated e.g. before a scheduled reset step. This is
 613        complicated by the fact that PME ranks are unaware of the
 614        reason why they were sent a pmerecvqxFINISH. To avoid
 615        communication deadlocks, we always do the communication for the
 616        report, even if we've decided not to write the report, because
 617        how long it takes to finish the run is not important when we've
 618        decided not to report on the simulation performance.
 619
 620        Further, we only report performance for dynamical integrators,
 621        because those are the only ones for which we plan to
 622        consider doing any optimizations. */
 623     bool printReport = EI_DYNAMICS(inputrec->eI) && SIMMASTER(cr);
 624
 625     if (printReport && !walltime_accounting_get_valid_finish(walltime_accounting))
 626     {
 627         GMX_LOG(mdlog.warning)
 628                 .asParagraph()
 629                 .appendText("Simulation ended prematurely, no performance report will be written.");
 630         printReport = false;
 631     }
 632
 633     t_nrnb*                 nrnb_tot;
 634     std::unique_ptr<t_nrnb> nrnbTotalStorage;
 635     if (cr->nnodes > 1)
 636     {
 637         nrnbTotalStorage = std::make_unique<t_nrnb>();
 638         nrnb_tot         = nrnbTotalStorage.get();
 639 #if GMX_MPI
 640         MPI_Allreduce(nrnb->n, nrnb_tot->n, eNRNB, MPI_DOUBLE, MPI_SUM, cr->mpi_comm_mysim);
 641 #endif
 642     }
 643     else
 644     {
 645         nrnb_tot = nrnb;
 646     }
 647
 648     elapsed_time = walltime_accounting_get_time_since_reset(walltime_accounting);
 649     elapsed_time_over_all_threads =
 650             walltime_accounting_get_time_since_reset_over_all_threads(walltime_accounting);
 651     if (cr->nnodes > 1)
 652     {
 653 #if GMX_MPI
 654         /* reduce elapsed_time over all MPI ranks in the current simulation */
 655         MPI_Allreduce(&elapsed_time, &elapsed_time_over_all_ranks, 1, MPI_DOUBLE, MPI_SUM, cr->mpi_comm_mysim);
 656         elapsed_time_over_all_ranks /= cr->nnodes;
 657         /* Reduce elapsed_time_over_all_threads over all MPI ranks in the
 658          * current simulation. */
 659         MPI_Allreduce(&elapsed_time_over_all_threads,
 660                       &elapsed_time_over_all_threads_over_all_ranks,
 661                       1,
 662                       MPI_DOUBLE,
 663                       MPI_SUM,
 664                       cr->mpi_comm_mysim);
 665 #endif
 666     }
 667     else
 668     {
 669         elapsed_time_over_all_ranks                  = elapsed_time;
 670         elapsed_time_over_all_threads_over_all_ranks = elapsed_time_over_all_threads;
 671     }
 672
 673     if (printReport)
 674     {
 675         print_flop(fplog, nrnb_tot, &nbfs, &mflop);
 676     }
 677
 678     if (thisRankHasDuty(cr, DUTY_PP) && DOMAINDECOMP(cr))
 679     {
 680         print_dd_statistics(cr, inputrec, fplog);
 681     }
 682
 683     /* TODO Move the responsibility for any scaling by thread counts
 684      * to the code that handled the thread region, so that there's a
 685      * mechanism to keep cycle counting working during the transition
 686      * to task parallelism. */
 687     int nthreads_pp  = gmx_omp_nthreads_get(emntNonbonded);
 688     int nthreads_pme = gmx_omp_nthreads_get(emntPME);
 689     wallcycle_scale_by_num_threads(
 690             wcycle, thisRankHasDuty(cr, DUTY_PME) && !thisRankHasDuty(cr, DUTY_PP), nthreads_pp, nthreads_pme);
 691     auto cycle_sum(wallcycle_sum(cr, wcycle));
 692
 693     if (printReport)
 694     {
 695         auto nbnxn_gpu_timings =
 696                 (nbv != nullptr && nbv->useGpu()) ? Nbnxm::gpu_get_timings(nbv->gpu_nbv) : nullptr;
 697         gmx_wallclock_gpu_pme_t pme_gpu_timings = {};
 698
 699         if (pme_gpu_task_enabled(pme))
 700         {
 701             pme_gpu_get_timings(pme, &pme_gpu_timings);
 702         }
 703         wallcycle_print(fplog,
 704                         mdlog,
 705                         cr->nnodes,
 706                         cr->npmenodes,
 707                         nthreads_pp,
 708                         nthreads_pme,
 709                         elapsed_time_over_all_ranks,
 710                         wcycle,
 711                         cycle_sum,
 712                         nbnxn_gpu_timings,
 713                         &pme_gpu_timings);
 714
 715         if (EI_DYNAMICS(inputrec->eI))
 716         {
 717             delta_t = inputrec->delta_t;
 718         }
 719
 720         if (fplog)
 721         {
 722             print_perf(fplog,
 723                        elapsed_time_over_all_threads_over_all_ranks,
 724                        elapsed_time_over_all_ranks,
 725                        walltime_accounting_get_nsteps_done_since_reset(walltime_accounting),
 726                        delta_t,
 727                        nbfs,
 728                        mflop);
 729         }
 730         if (bWriteStat)
 731         {
 732             print_perf(stderr,
 733                        elapsed_time_over_all_threads_over_all_ranks,
 734                        elapsed_time_over_all_ranks,
 735                        walltime_accounting_get_nsteps_done_since_reset(walltime_accounting),
 736                        delta_t,
 737                        nbfs,
 738                        mflop);
 739         }
 740     }
 741 }
 742
 743 int Mdrunner::mdrunner()
 744 {
 745     matrix                    box;
 746     t_forcerec*               fr               = nullptr;
 747     real                      ewaldcoeff_q     = 0;
 748     real                      ewaldcoeff_lj    = 0;
 749     int                       nChargePerturbed = -1, nTypePerturbed = 0;
 750     gmx_wallcycle_t           wcycle;
 751     gmx_walltime_accounting_t walltime_accounting = nullptr;
 752     MembedHolder              membedHolder(filenames.size(), filenames.data());
 753
 754     /* CAUTION: threads may be started later on in this function, so
 755        cr doesn't reflect the final parallel state right now */
 756     gmx_mtop_t mtop;
 757
 758     /* TODO: inputrec should tell us whether we use an algorithm, not a file option */
 759     const bool doEssentialDynamics = opt2bSet("-ei", filenames.size(), filenames.data());
 760     const bool doRerun             = mdrunOptions.rerun;
 761
 762     // Handle task-assignment related user options.
 763     EmulateGpuNonbonded emulateGpuNonbonded =
 764             (getenv("GMX_EMULATE_GPU") != nullptr ? EmulateGpuNonbonded::Yes : EmulateGpuNonbonded::No);
 765
 766     std::vector<int> userGpuTaskAssignment;
 767     try
 768     {
 769         userGpuTaskAssignment = parseUserTaskAssignmentString(hw_opt.userGpuTaskAssignment);
 770     }
 771     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
 772     auto nonbondedTarget = findTaskTarget(nbpu_opt);
 773     auto pmeTarget       = findTaskTarget(pme_opt);
 774     auto pmeFftTarget    = findTaskTarget(pme_fft_opt);
 775     auto bondedTarget    = findTaskTarget(bonded_opt);
 776     auto updateTarget    = findTaskTarget(update_opt);
 777
 778     FILE* fplog = nullptr;
 779     // If we are appending, we don't write log output because we need
 780     // to check that the old log file matches what the checkpoint file
 781     // expects. Otherwise, we should start to write log output now if
 782     // there is a file ready for it.
 783     if (logFileHandle != nullptr && startingBehavior != StartingBehavior::RestartWithAppending)
 784     {
 785         fplog = gmx_fio_getfp(logFileHandle);
 786     }
 787     const bool isSimulationMasterRank = findIsSimulationMasterRank(ms, simulationCommunicator);
 788     gmx::LoggerOwner logOwner(buildLogger(fplog, isSimulationMasterRank));
 789     gmx::MDLogger    mdlog(logOwner.logger());
 790
 791     gmx_print_detected_hardware(fplog, isSimulationMasterRank && isMasterSim(ms), mdlog, hwinfo_);
 792
 793     std::vector<int> gpuIdsToUse = makeGpuIdsToUse(hwinfo_->deviceInfoList, hw_opt.gpuIdsAvailable);
 794     const int        numDevicesToUse = gmx::ssize(gpuIdsToUse);
 795
 796     // Print citation requests after all software/hardware printing
 797     pleaseCiteGromacs(fplog);
 798
 799     // Note: legacy program logic relies on checking whether these pointers are assigned.
 800     // Objects may or may not be allocated later.
 801     std::unique_ptr<t_inputrec> inputrec;
 802     std::unique_ptr<t_state>    globalState;
 803
 804     auto partialDeserializedTpr = std::make_unique<PartialDeserializedTprFile>();
 805
 806     if (isSimulationMasterRank)
 807     {
 808         // Allocate objects to be initialized by later function calls.
 809         /* Only the master rank has the global state */
 810         globalState = std::make_unique<t_state>();
 811         inputrec    = std::make_unique<t_inputrec>();
 812
 813         /* Read (nearly) all data required for the simulation
 814          * and keep the partly serialized tpr contents to send to other ranks later
 815          */
 816         applyGlobalSimulationState(
 817                 *inputHolder_.get(), partialDeserializedTpr.get(), globalState.get(), inputrec.get(), &mtop);
 818     }
 819
 820     /* Check and update the hardware options for internal consistency */
 821     checkAndUpdateHardwareOptions(
 822             mdlog, &hw_opt, isSimulationMasterRank, domdecOptions.numPmeRanks, inputrec.get());
 823
 824     if (GMX_THREAD_MPI && isSimulationMasterRank)
 825     {
 826         bool useGpuForNonbonded = false;
 827         bool useGpuForPme       = false;
 828         try
 829         {
 830             GMX_RELEASE_ASSERT(inputrec != nullptr, "Keep the compiler happy");
 831
 832             // If the user specified the number of ranks, then we must
 833             // respect that, but in default mode, we need to allow for
 834             // the number of GPUs to choose the number of ranks.
 835             auto canUseGpuForNonbonded = buildSupportsNonbondedOnGpu(nullptr);
 836             useGpuForNonbonded         = decideWhetherToUseGpusForNonbondedWithThreadMpi(
 837                     nonbondedTarget,
 838                     numDevicesToUse,
 839                     userGpuTaskAssignment,
 840                     emulateGpuNonbonded,
 841                     canUseGpuForNonbonded,
 842                     gpuAccelerationOfNonbondedIsUseful(mdlog, *inputrec, GMX_THREAD_MPI),
 843                     hw_opt.nthreads_tmpi);
 844             useGpuForPme = decideWhetherToUseGpusForPmeWithThreadMpi(useGpuForNonbonded,
 845                                                                      pmeTarget,
 846                                                                      numDevicesToUse,
 847                                                                      userGpuTaskAssignment,
 848                                                                      *hwinfo_,
 849                                                                      *inputrec,
 850                                                                      hw_opt.nthreads_tmpi,
 851                                                                      domdecOptions.numPmeRanks);
 852         }
 853         GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
 854
 855         /* Determine how many thread-MPI ranks to start.
 856          *
 857          * TODO Over-writing the user-supplied value here does
 858          * prevent any possible subsequent checks from working
 859          * correctly. */
 860         hw_opt.nthreads_tmpi = get_nthreads_mpi(hwinfo_,
 861                                                 &hw_opt,
 862                                                 numDevicesToUse,
 863                                                 useGpuForNonbonded,
 864                                                 useGpuForPme,
 865                                                 inputrec.get(),
 866                                                 &mtop,
 867                                                 mdlog,
 868                                                 membedHolder.doMembed());
 869
 870         // Now start the threads for thread MPI.
 871         spawnThreads(hw_opt.nthreads_tmpi);
 872         // The spawned threads enter mdrunner() and execution of
 873         // master and spawned threads joins at the end of this block.
 874     }
 875
 876     GMX_RELEASE_ASSERT(ms || simulationCommunicator != MPI_COMM_NULL,
 877                        "Must have valid communicator unless running a multi-simulation");
 878     CommrecHandle crHandle = init_commrec(simulationCommunicator);
 879     t_commrec*    cr       = crHandle.get();
 880     GMX_RELEASE_ASSERT(cr != nullptr, "Must have valid commrec");
 881
 882     PhysicalNodeCommunicator physicalNodeComm(libraryWorldCommunicator, gmx_physicalnode_id_hash());
 883
 884     // If we detected the topology on this system, double-check that it makes sense
 885     if (hwinfo_->hardwareTopology->isThisSystem())
 886     {
 887         hardwareTopologyDoubleCheckDetection(mdlog, *hwinfo_->hardwareTopology);
 888     }
 889
 890     if (PAR(cr))
 891     {
 892         /* now broadcast everything to the non-master nodes/threads: */
 893         if (!isSimulationMasterRank)
 894         {
 895             // Until now, only the master rank has a non-null pointer.
 896             // On non-master ranks, allocate the object that will receive data in the following call.
 897             inputrec = std::make_unique<t_inputrec>();
 898         }
 899         init_parallel(cr->mpiDefaultCommunicator,
 900                       MASTER(cr),
 901                       inputrec.get(),
 902                       &mtop,
 903                       partialDeserializedTpr.get());
 904     }
 905     GMX_RELEASE_ASSERT(inputrec != nullptr, "All ranks should have a valid inputrec now");
 906     partialDeserializedTpr.reset(nullptr);
 907
 908     GMX_RELEASE_ASSERT(
 909             !inputrec->useConstantAcceleration,
 910             "Linear acceleration has been removed in GROMACS 2022, and was broken for many years "
 911             "before that. Use GROMACS 4.5 or earlier if you need this feature.");
 912
 913     // Now the number of ranks is known to all ranks, and each knows
 914     // the inputrec read by the master rank. The ranks can now all run
 915     // the task-deciding functions and will agree on the result
 916     // without needing to communicate.
 917     const bool useDomainDecomposition = (PAR(cr) && !(EI_TPI(inputrec->eI) || inputrec->eI == eiNM));
 918
 919     // Note that these variables describe only their own node.
 920     //
 921     // Note that when bonded interactions run on a GPU they always run
 922     // alongside a nonbonded task, so do not influence task assignment
 923     // even though they affect the force calculation workload.
 924     bool useGpuForNonbonded = false;
 925     bool useGpuForPme       = false;
 926     bool useGpuForBonded    = false;
 927     bool useGpuForUpdate    = false;
 928     bool gpusWereDetected   = hwinfo_->ngpu_compatible_tot > 0;
 929     try
 930     {
 931         // It's possible that there are different numbers of GPUs on
 932         // different nodes, which is the user's responsibility to
 933         // handle. If unsuitable, we will notice that during task
 934         // assignment.
 935         auto canUseGpuForNonbonded = buildSupportsNonbondedOnGpu(nullptr);
 936         useGpuForNonbonded         = decideWhetherToUseGpusForNonbonded(
 937                 nonbondedTarget,
 938                 userGpuTaskAssignment,
 939                 emulateGpuNonbonded,
 940                 canUseGpuForNonbonded,
 941                 gpuAccelerationOfNonbondedIsUseful(mdlog, *inputrec, !GMX_THREAD_MPI),
 942                 gpusWereDetected);
 943         useGpuForPme    = decideWhetherToUseGpusForPme(useGpuForNonbonded,
 944                                                     pmeTarget,
 945                                                     userGpuTaskAssignment,
 946                                                     *hwinfo_,
 947                                                     *inputrec,
 948                                                     cr->sizeOfDefaultCommunicator,
 949                                                     domdecOptions.numPmeRanks,
 950                                                     gpusWereDetected);
 951         useGpuForBonded = decideWhetherToUseGpusForBonded(
 952                 useGpuForNonbonded, useGpuForPme, bondedTarget, *inputrec, mtop, domdecOptions.numPmeRanks, gpusWereDetected);
 953     }
 954     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
 955
 956     const PmeRunMode pmeRunMode = determinePmeRunMode(useGpuForPme, pmeFftTarget, *inputrec);
 957
 958     // Initialize development feature flags that enabled by environment variable
 959     // and report those features that are enabled.
 960     const DevelopmentFeatureFlags devFlags =
 961             manageDevelopmentFeatures(mdlog, useGpuForNonbonded, pmeRunMode);
 962
 963     const bool useModularSimulator = checkUseModularSimulator(false,
 964                                                               inputrec.get(),
 965                                                               doRerun,
 966                                                               mtop,
 967                                                               ms,
 968                                                               replExParams,
 969                                                               nullptr,
 970                                                               doEssentialDynamics,
 971                                                               membedHolder.doMembed());
 972
 973     // Build restraints.
 974     // TODO: hide restraint implementation details from Mdrunner.
 975     // There is nothing unique about restraints at this point as far as the
 976     // Mdrunner is concerned. The Mdrunner should just be getting a sequence of
 977     // factory functions from the SimulationContext on which to call mdModules_->add().
 978     // TODO: capture all restraints into a single RestraintModule, passed to the runner builder.
 979     for (auto&& restraint : restraintManager_->getRestraints())
 980     {
 981         auto module = RestraintMDModule::create(restraint, restraint->sites());
 982         mdModules_->add(std::move(module));
 983     }
 984
 985     // TODO: Error handling
 986     mdModules_->assignOptionsToModules(*inputrec->params, nullptr);
 987     // now that the MdModules know their options, they know which callbacks to sign up to
 988     mdModules_->subscribeToSimulationSetupNotifications();
 989     const auto& mdModulesNotifier = mdModules_->notifier().simulationSetupNotifications_;
 990
 991     if (inputrec->internalParameters != nullptr)
 992     {
 993         mdModulesNotifier.notify(*inputrec->internalParameters);
 994     }
 995
 996     if (fplog != nullptr)
 997     {
 998         pr_inputrec(fplog, 0, "Input Parameters", inputrec.get(), FALSE);
 999         fprintf(fplog, "\n");
1000     }
1001
1002     if (SIMMASTER(cr))
1003     {
1004         /* In rerun, set velocities to zero if present */
1005         if (doRerun && ((globalState->flags & (1 << estV)) != 0))
1006         {
1007             // rerun does not use velocities
1008             GMX_LOG(mdlog.info)
1009                     .asParagraph()
1010                     .appendText(
1011                             "Rerun trajectory contains velocities. Rerun does only evaluate "
1012                             "potential energy and forces. The velocities will be ignored.");
1013             for (int i = 0; i < globalState->natoms; i++)
1014             {
1015                 clear_rvec(globalState->v[i]);
1016             }
1017             globalState->flags &= ~(1 << estV);
1018         }
1019
1020         /* now make sure the state is initialized and propagated */
1021         set_state_entries(globalState.get(), inputrec.get(), useModularSimulator);
1022     }
1023
1024     /* NM and TPI parallelize over force/energy calculations, not atoms,
1025      * so we need to initialize and broadcast the global state.
1026      */
1027     if (inputrec->eI == eiNM || inputrec->eI == eiTPI)
1028     {
1029         if (!MASTER(cr))
1030         {
1031             globalState = std::make_unique<t_state>();
1032         }
1033         broadcastStateWithoutDynamics(
1034                 cr->mpiDefaultCommunicator, DOMAINDECOMP(cr), PAR(cr), globalState.get());
1035     }
1036
1037     /* A parallel command line option consistency check that we can
1038        only do after any threads have started. */
1039     if (!PAR(cr)
1040         && (domdecOptions.numCells[XX] > 1 || domdecOptions.numCells[YY] > 1
1041             || domdecOptions.numCells[ZZ] > 1 || domdecOptions.numPmeRanks > 0))
1042     {
1043         gmx_fatal(FARGS,
1044                   "The -dd or -npme option request a parallel simulation, "
1045 #if !GMX_MPI
1046                   "but %s was compiled without threads or MPI enabled",
1047                   output_env_get_program_display_name(oenv));
1048 #elif GMX_THREAD_MPI
1049                   "but the number of MPI-threads (option -ntmpi) is not set or is 1");
1050 #else
1051                   "but %s was not started through mpirun/mpiexec or only one rank was requested "
1052                   "through mpirun/mpiexec",
1053                   output_env_get_program_display_name(oenv));
1054 #endif
1055     }
1056
1057     if (doRerun && (EI_ENERGY_MINIMIZATION(inputrec->eI) || eiNM == inputrec->eI))
1058     {
1059         gmx_fatal(FARGS,
1060                   "The .mdp file specified an energy mininization or normal mode algorithm, and "
1061                   "these are not compatible with mdrun -rerun");
1062     }
1063
1064     if (!(EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype)))
1065     {
1066         if (domdecOptions.numPmeRanks > 0)
1067         {
1068             gmx_fatal_collective(FARGS,
1069                                  cr->mpiDefaultCommunicator,
1070                                  MASTER(cr),
1071                                  "PME-only ranks are requested, but the system does not use PME "
1072                                  "for electrostatics or LJ");
1073         }
1074
1075         domdecOptions.numPmeRanks = 0;
1076     }
1077
1078     if (useGpuForNonbonded && domdecOptions.numPmeRanks < 0)
1079     {
1080         /* With NB GPUs we don't automatically use PME-only CPU ranks. PME ranks can
1081          * improve performance with many threads per GPU, since our OpenMP
1082          * scaling is bad, but it's difficult to automate the setup.
1083          */
1084         domdecOptions.numPmeRanks = 0;
1085     }
1086     if (useGpuForPme)
1087     {
1088         if (domdecOptions.numPmeRanks < 0)
1089         {
1090             domdecOptions.numPmeRanks = 0;
1091             // TODO possibly print a note that one can opt-in for a separate PME GPU rank?
1092         }
1093         else
1094         {
1095             GMX_RELEASE_ASSERT(domdecOptions.numPmeRanks <= 1,
1096                                "PME GPU decomposition is not supported");
1097         }
1098     }
1099
1100     /* NMR restraints must be initialized before load_checkpoint,
1101      * since with time averaging the history is added to t_state.
1102      * For proper consistency check we therefore need to extend
1103      * t_state here.
1104      * So the PME-only nodes (if present) will also initialize
1105      * the distance restraints.
1106      */
1107
1108     /* This needs to be called before read_checkpoint to extend the state */
1109     t_disresdata* disresdata;
1110     snew(disresdata, 1);
1111     init_disres(fplog,
1112                 &mtop,
1113                 inputrec.get(),
1114                 DisResRunMode::MDRun,
1115                 MASTER(cr) ? DDRole::Master : DDRole::Agent,
1116                 PAR(cr) ? NumRanks::Multiple : NumRanks::Single,
1117                 cr->mpi_comm_mysim,
1118                 ms,
1119                 disresdata,
1120                 globalState.get(),
1121                 replExParams.exchangeInterval > 0);
1122
1123     t_oriresdata* oriresdata;
1124     snew(oriresdata, 1);
1125     init_orires(fplog, &mtop, inputrec.get(), cr, ms, globalState.get(), oriresdata);
1126
1127     auto deform = prepareBoxDeformation(globalState != nullptr ? globalState->box : box,
1128                                         MASTER(cr) ? DDRole::Master : DDRole::Agent,
1129                                         PAR(cr) ? NumRanks::Multiple : NumRanks::Single,
1130                                         cr->mpi_comm_mygroup,
1131                                         *inputrec);
1132
1133 #if GMX_FAHCORE
1134     /* We have to remember the generation's first step before reading checkpoint.
1135        This way, we can report to the F@H core both the generation's first step
1136        and the restored first step, thus making it able to distinguish between
1137        an interruption/resume and start of the n-th generation simulation.
1138        Having this information, the F@H core can correctly calculate and report
1139        the progress.
1140      */
1141     int gen_first_step = 0;
1142     if (MASTER(cr))
1143     {
1144         gen_first_step = inputrec->init_step;
1145     }
1146 #endif
1147
1148     ObservablesHistory observablesHistory = {};
1149
1150     auto modularSimulatorCheckpointData = std::make_unique<ReadCheckpointDataHolder>();
1151     if (startingBehavior != StartingBehavior::NewSimulation)
1152     {
1153         /* Check if checkpoint file exists before doing continuation.
1154          * This way we can use identical input options for the first and subsequent runs...
1155          */
1156         if (mdrunOptions.numStepsCommandline > -2)
1157         {
1158             /* Temporarily set the number of steps to unlimited to avoid
1159              * triggering the nsteps check in load_checkpoint().
1160              * This hack will go away soon when the -nsteps option is removed.
1161              */
1162             inputrec->nsteps = -1;
1163         }
1164
1165         // Finish applying initial simulation state information from external sources on all ranks.
1166         // Reconcile checkpoint file data with Mdrunner state established up to this point.
1167         applyLocalState(*inputHolder_.get(),
1168                         logFileHandle,
1169                         cr,
1170                         domdecOptions.numCells,
1171                         inputrec.get(),
1172                         globalState.get(),
1173                         &observablesHistory,
1174                         mdrunOptions.reproducible,
1175                         mdModules_->notifier(),
1176                         modularSimulatorCheckpointData.get(),
1177                         useModularSimulator);
1178         // TODO: (#3652) Synchronize filesystem state, SimulationInput contents, and program
1179         //  invariants
1180         //  on all code paths.
1181         // Write checkpoint or provide hook to update SimulationInput.
1182         // If there was a checkpoint file, SimulationInput contains more information
1183         // than if there wasn't. At this point, we have synchronized the in-memory
1184         // state with the filesystem state only for restarted simulations. We should
1185         // be calling applyLocalState unconditionally and expect that the completeness
1186         // of SimulationInput is not dependent on its creation method.
1187
1188         if (startingBehavior == StartingBehavior::RestartWithAppending && logFileHandle)
1189         {
1190             // Now we can start normal logging to the truncated log file.
1191             fplog = gmx_fio_getfp(logFileHandle);
1192             prepareLogAppending(fplog);
1193             logOwner = buildLogger(fplog, MASTER(cr));
1194             mdlog    = logOwner.logger();
1195         }
1196     }
1197
1198 #if GMX_FAHCORE
1199     if (MASTER(cr))
1200     {
1201         fcRegisterSteps(inputrec->nsteps + inputrec->init_step, gen_first_step);
1202     }
1203 #endif
1204
1205     if (mdrunOptions.numStepsCommandline > -2)
1206     {
1207         GMX_LOG(mdlog.info)
1208                 .asParagraph()
1209                 .appendText(
1210                         "The -nsteps functionality is deprecated, and may be removed in a future "
1211                         "version. "
1212                         "Consider using gmx convert-tpr -nsteps or changing the appropriate .mdp "
1213                         "file field.");
1214     }
1215     /* override nsteps with value set on the commandline */
1216     override_nsteps_cmdline(mdlog, mdrunOptions.numStepsCommandline, inputrec.get());
1217
1218     if (isSimulationMasterRank)
1219     {
1220         copy_mat(globalState->box, box);
1221     }
1222
1223     if (PAR(cr))
1224     {
1225         gmx_bcast(sizeof(box), box, cr->mpiDefaultCommunicator);
1226     }
1227
1228     if (inputrec->cutoff_scheme != ecutsVERLET)
1229     {
1230         gmx_fatal(FARGS,
1231                   "This group-scheme .tpr file can no longer be run by mdrun. Please update to the "
1232                   "Verlet scheme, or use an earlier version of GROMACS if necessary.");
1233     }
1234     /* Update rlist and nstlist. */
1235     /* Note: prepare_verlet_scheme is calling increaseNstlist(...), which (while attempting to
1236      * increase rlist) tries to check if the newly chosen value fits with the DD scheme. As this is
1237      * run before any DD scheme is set up, this check is never executed. See #3334 for more details.
1238      */
1239     prepare_verlet_scheme(fplog,
1240                           cr,
1241                           inputrec.get(),
1242                           nstlist_cmdline,
1243                           &mtop,
1244                           box,
1245                           useGpuForNonbonded || (emulateGpuNonbonded == EmulateGpuNonbonded::Yes),
1246                           *hwinfo_->cpuInfo);
1247
1248     // This builder is necessary while we have multi-part construction
1249     // of DD. Before DD is constructed, we use the existence of
1250     // the builder object to indicate that further construction of DD
1251     // is needed.
1252     std::unique_ptr<DomainDecompositionBuilder> ddBuilder;
1253     if (useDomainDecomposition)
1254     {
1255         ddBuilder = std::make_unique<DomainDecompositionBuilder>(
1256                 mdlog,
1257                 cr,
1258                 domdecOptions,
1259                 mdrunOptions,
1260                 mtop,
1261                 *inputrec,
1262                 box,
1263                 positionsFromStatePointer(globalState.get()));
1264     }
1265     else
1266     {
1267         /* PME, if used, is done on all nodes with 1D decomposition */
1268         cr->nnodes     = cr->sizeOfDefaultCommunicator;
1269         cr->sim_nodeid = cr->rankInDefaultCommunicator;
1270         cr->nodeid     = cr->rankInDefaultCommunicator;
1271         cr->npmenodes  = 0;
1272         cr->duty       = (DUTY_PP | DUTY_PME);
1273
1274         if (inputrec->pbcType == PbcType::Screw)
1275         {
1276             gmx_fatal(FARGS, "pbc=screw is only implemented with domain decomposition");
1277         }
1278     }
1279
1280     // Produce the task assignment for this rank - done after DD is constructed
1281     GpuTaskAssignments gpuTaskAssignments = GpuTaskAssignmentsBuilder::build(
1282             gpuIdsToUse,
1283             userGpuTaskAssignment,
1284             *hwinfo_,
1285             simulationCommunicator,
1286             physicalNodeComm,
1287             nonbondedTarget,
1288             pmeTarget,
1289             bondedTarget,
1290             updateTarget,
1291             useGpuForNonbonded,
1292             useGpuForPme,
1293             thisRankHasDuty(cr, DUTY_PP),
1294             // TODO cr->duty & DUTY_PME should imply that a PME
1295             // algorithm is active, but currently does not.
1296             EEL_PME(inputrec->coulombtype) && thisRankHasDuty(cr, DUTY_PME));
1297
1298     // Get the device handles for the modules, nullptr when no task is assigned.
1299     int                deviceId   = -1;
1300     DeviceInformation* deviceInfo = gpuTaskAssignments.initDevice(&deviceId);
1301
1302     // timing enabling - TODO put this in gpu_utils (even though generally this is just option handling?)
1303     bool useTiming = true;
1304
1305     if (GMX_GPU_CUDA)
1306     {
1307         /* WARNING: CUDA timings are incorrect with multiple streams.
1308          *          This is the main reason why they are disabled by default.
1309          */
1310         // TODO: Consider turning on by default when we can detect nr of streams.
1311         useTiming = (getenv("GMX_ENABLE_GPU_TIMING") != nullptr);
1312     }
1313     else if (GMX_GPU_OPENCL)
1314     {
1315         useTiming = (getenv("GMX_DISABLE_GPU_TIMING") == nullptr);
1316     }
1317
1318     // TODO Currently this is always built, yet DD partition code
1319     // checks if it is built before using it. Probably it should
1320     // become an MDModule that is made only when another module
1321     // requires it (e.g. pull, CompEl, density fitting), so that we
1322     // don't update the local atom sets unilaterally every step.
1323     LocalAtomSetManager atomSets;
1324     if (ddBuilder)
1325     {
1326         // TODO Pass the GPU streams to ddBuilder to use in buffer
1327         // transfers (e.g. halo exchange)
1328         cr->dd = ddBuilder->build(&atomSets);
1329         // The builder's job is done, so destruct it
1330         ddBuilder.reset(nullptr);
1331         // Note that local state still does not exist yet.
1332     }
1333
1334     // The GPU update is decided here because we need to know whether the constraints or
1335     // SETTLEs can span accross the domain borders (i.e. whether or not update groups are
1336     // defined). This is only known after DD is initialized, hence decision on using GPU
1337     // update is done so late.
1338     try
1339     {
1340         const bool useUpdateGroups = cr->dd ? ddUsesUpdateGroups(*cr->dd) : false;
1341
1342         useGpuForUpdate = decideWhetherToUseGpuForUpdate(useDomainDecomposition,
1343                                                          useUpdateGroups,
1344                                                          pmeRunMode,
1345                                                          domdecOptions.numPmeRanks > 0,
1346                                                          useGpuForNonbonded,
1347                                                          updateTarget,
1348                                                          gpusWereDetected,
1349                                                          *inputrec,
1350                                                          mtop,
1351                                                          doEssentialDynamics,
1352                                                          gmx_mtop_ftype_count(mtop, F_ORIRES) > 0,
1353                                                          replExParams.exchangeInterval > 0,
1354                                                          doRerun,
1355                                                          devFlags,
1356                                                          mdlog);
1357     }
1358     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1359
1360     const bool printHostName = (cr->nnodes > 1);
1361     gpuTaskAssignments.reportGpuUsage(mdlog, printHostName, useGpuForBonded, pmeRunMode, useGpuForUpdate);
1362
1363     const bool disableNonbondedCalculation = (getenv("GMX_NO_NONBONDED") != nullptr);
1364     if (disableNonbondedCalculation)
1365     {
1366         /* turn off non-bonded calculations */
1367         GMX_LOG(mdlog.warning)
1368                 .asParagraph()
1369                 .appendText(
1370                         "Found environment variable GMX_NO_NONBONDED.\n"
1371                         "Disabling nonbonded calculations.");
1372     }
1373
1374     MdrunScheduleWorkload runScheduleWork;
1375
1376     bool useGpuDirectHalo = decideWhetherToUseGpuForHalo(devFlags,
1377                                                          havePPDomainDecomposition(cr),
1378                                                          useGpuForNonbonded,
1379                                                          useModularSimulator,
1380                                                          doRerun,
1381                                                          EI_ENERGY_MINIMIZATION(inputrec->eI));
1382
1383     // Also populates the simulation constant workload description.
1384     runScheduleWork.simulationWork = createSimulationWorkload(*inputrec,
1385                                                               disableNonbondedCalculation,
1386                                                               devFlags,
1387                                                               useGpuForNonbonded,
1388                                                               pmeRunMode,
1389                                                               useGpuForBonded,
1390                                                               useGpuForUpdate,
1391                                                               useGpuDirectHalo);
1392
1393     std::unique_ptr<DeviceStreamManager> deviceStreamManager = nullptr;
1394
1395     if (deviceInfo != nullptr)
1396     {
1397         if (DOMAINDECOMP(cr) && thisRankHasDuty(cr, DUTY_PP))
1398         {
1399             dd_setup_dlb_resource_sharing(cr, deviceId);
1400         }
1401         deviceStreamManager = std::make_unique<DeviceStreamManager>(
1402                 *deviceInfo, havePPDomainDecomposition(cr), runScheduleWork.simulationWork, useTiming);
1403     }
1404
1405     // If the user chose a task assignment, give them some hints
1406     // where appropriate.
1407     if (!userGpuTaskAssignment.empty())
1408     {
1409         gpuTaskAssignments.logPerformanceHints(mdlog, numDevicesToUse);
1410     }
1411
1412     if (PAR(cr))
1413     {
1414         /* After possible communicator splitting in make_dd_communicators.
1415          * we can set up the intra/inter node communication.
1416          */
1417         gmx_setup_nodecomm(fplog, cr);
1418     }
1419
1420 #if GMX_MPI
1421     if (isMultiSim(ms))
1422     {
1423         GMX_LOG(mdlog.warning)
1424                 .asParagraph()
1425                 .appendTextFormatted(
1426                         "This is simulation %d out of %d running as a composite GROMACS\n"
1427                         "multi-simulation job. Setup for this simulation:\n",
1428                         ms->simulationIndex_,
1429                         ms->numSimulations_);
1430     }
1431     GMX_LOG(mdlog.warning)
1432             .appendTextFormatted("Using %d MPI %s\n",
1433                                  cr->nnodes,
1434 #    if GMX_THREAD_MPI
1435                                  cr->nnodes == 1 ? "thread" : "threads"
1436 #    else
1437                                  cr->nnodes == 1 ? "process" : "processes"
1438 #    endif
1439             );
1440     fflush(stderr);
1441 #endif
1442
1443     // If mdrun -pin auto honors any affinity setting that already
1444     // exists. If so, it is nice to provide feedback about whether
1445     // that existing affinity setting was from OpenMP or something
1446     // else, so we run this code both before and after we initialize
1447     // the OpenMP support.
1448     gmx_check_thread_affinity_set(mdlog, &hw_opt, hwinfo_->nthreads_hw_avail, FALSE);
1449     /* Check and update the number of OpenMP threads requested */
1450     checkAndUpdateRequestedNumOpenmpThreads(
1451             &hw_opt, *hwinfo_, cr, ms, physicalNodeComm.size_, pmeRunMode, mtop, *inputrec);
1452
1453     gmx_omp_nthreads_init(mdlog,
1454                           cr,
1455                           hwinfo_->nthreads_hw_avail,
1456                           physicalNodeComm.size_,
1457                           hw_opt.nthreads_omp,
1458                           hw_opt.nthreads_omp_pme,
1459                           !thisRankHasDuty(cr, DUTY_PP));
1460
1461     // Enable FP exception detection, but not in
1462     // Release mode and not for compilers with known buggy FP
1463     // exception support (clang with any optimization) or suspected
1464     // buggy FP exception support (gcc 7.* with optimization).
1465 #if !defined NDEBUG                                                                         \
1466         && !((defined __clang__ || (defined(__GNUC__) && !defined(__ICC) && __GNUC__ == 7)) \
1467              && defined __OPTIMIZE__)
1468     const bool bEnableFPE = true;
1469 #else
1470     const bool bEnableFPE = false;
1471 #endif
1472     // FIXME - reconcile with gmx_feenableexcept() call from CommandLineModuleManager::run()
1473     if (bEnableFPE)
1474     {
1475         gmx_feenableexcept();
1476     }
1477
1478     /* Now that we know the setup is consistent, check for efficiency */
1479     check_resource_division_efficiency(
1480             hwinfo_, gpuTaskAssignments.thisRankHasAnyGpuTask(), mdrunOptions.ntompOptionIsSet, cr, mdlog);
1481
1482     /* getting number of PP/PME threads on this MPI / tMPI rank.
1483        PME: env variable should be read only on one node to make sure it is
1484        identical everywhere;
1485      */
1486     const int numThreadsOnThisRank = thisRankHasDuty(cr, DUTY_PP) ? gmx_omp_nthreads_get(emntNonbonded)
1487                                                                   : gmx_omp_nthreads_get(emntPME);
1488     checkHardwareOversubscription(
1489             numThreadsOnThisRank, cr->nodeid, *hwinfo_->hardwareTopology, physicalNodeComm, mdlog);
1490
1491     // Enable Peer access between GPUs where available
1492     // Only for DD, only master PP rank needs to perform setup, and only if thread MPI plus
1493     // any of the GPU communication features are active.
1494     if (DOMAINDECOMP(cr) && MASTER(cr) && thisRankHasDuty(cr, DUTY_PP) && GMX_THREAD_MPI
1495         && (runScheduleWork.simulationWork.useGpuHaloExchange
1496             || runScheduleWork.simulationWork.useGpuPmePpCommunication))
1497     {
1498         setupGpuDevicePeerAccess(gpuIdsToUse, mdlog);
1499     }
1500
1501     if (hw_opt.threadAffinity != ThreadAffinity::Off)
1502     {
1503         /* Before setting affinity, check whether the affinity has changed
1504          * - which indicates that probably the OpenMP library has changed it
1505          * since we first checked).
1506          */
1507         gmx_check_thread_affinity_set(mdlog, &hw_opt, hwinfo_->nthreads_hw_avail, TRUE);
1508
1509         int numThreadsOnThisNode, intraNodeThreadOffset;
1510         analyzeThreadsOnThisNode(
1511                 physicalNodeComm, numThreadsOnThisRank, &numThreadsOnThisNode, &intraNodeThreadOffset);
1512
1513         /* Set the CPU affinity */
1514         gmx_set_thread_affinity(mdlog,
1515                                 cr,
1516                                 &hw_opt,
1517                                 *hwinfo_->hardwareTopology,
1518                                 numThreadsOnThisRank,
1519                                 numThreadsOnThisNode,
1520                                 intraNodeThreadOffset,
1521                                 nullptr);
1522     }
1523
1524     if (mdrunOptions.timingOptions.resetStep > -1)
1525     {
1526         GMX_LOG(mdlog.info)
1527                 .asParagraph()
1528                 .appendText(
1529                         "The -resetstep functionality is deprecated, and may be removed in a "
1530                         "future version.");
1531     }
1532     wcycle = wallcycle_init(fplog, mdrunOptions.timingOptions.resetStep, cr);
1533
1534     if (PAR(cr))
1535     {
1536         /* Master synchronizes its value of reset_counters with all nodes
1537          * including PME only nodes */
1538         int64_t reset_counters = wcycle_get_reset_counters(wcycle);
1539         gmx_bcast(sizeof(reset_counters), &reset_counters, cr->mpi_comm_mysim);
1540         wcycle_set_reset_counters(wcycle, reset_counters);
1541     }
1542
1543     // Membrane embedding must be initialized before we call init_forcerec()
1544     membedHolder.initializeMembed(fplog,
1545                                   filenames.size(),
1546                                   filenames.data(),
1547                                   &mtop,
1548                                   inputrec.get(),
1549                                   globalState.get(),
1550                                   cr,
1551                                   &mdrunOptions.checkpointOptions.period);
1552
1553     const bool               thisRankHasPmeGpuTask = gpuTaskAssignments.thisRankHasPmeGpuTask();
1554     std::unique_ptr<MDAtoms> mdAtoms;
1555     std::unique_ptr<VirtualSitesHandler> vsite;
1556     std::unique_ptr<GpuBonded>           gpuBonded;
1557
1558     t_nrnb nrnb;
1559     if (thisRankHasDuty(cr, DUTY_PP))
1560     {
1561         mdModulesNotifier.notify(*cr);
1562         mdModulesNotifier.notify(&atomSets);
1563         mdModulesNotifier.notify(inputrec->pbcType);
1564         mdModulesNotifier.notify(SimulationTimeStep{ inputrec->delta_t });
1565         /* Initiate forcerecord */
1566         fr                 = new t_forcerec;
1567         fr->forceProviders = mdModules_->initForceProviders();
1568         init_forcerec(fplog,
1569                       mdlog,
1570                       fr,
1571                       inputrec.get(),
1572                       &mtop,
1573                       cr,
1574                       box,
1575                       opt2fn("-table", filenames.size(), filenames.data()),
1576                       opt2fn("-tablep", filenames.size(), filenames.data()),
1577                       opt2fns("-tableb", filenames.size(), filenames.data()),
1578                       pforce);
1579         // Dirty hack, for fixing disres and orires should be made mdmodules
1580         fr->fcdata->disres = disresdata;
1581         fr->fcdata->orires = oriresdata;
1582
1583         // Save a handle to device stream manager to use elsewhere in the code
1584         // TODO: Forcerec is not a correct place to store it.
1585         fr->deviceStreamManager = deviceStreamManager.get();
1586
1587         if (runScheduleWork.simulationWork.useGpuPmePpCommunication && !thisRankHasDuty(cr, DUTY_PME))
1588         {
1589             GMX_RELEASE_ASSERT(
1590                     deviceStreamManager != nullptr,
1591                     "GPU device stream manager should be valid in order to use PME-PP direct "
1592                     "communications.");
1593             GMX_RELEASE_ASSERT(
1594                     deviceStreamManager->streamIsValid(DeviceStreamType::PmePpTransfer),
1595                     "GPU PP-PME stream should be valid in order to use GPU PME-PP direct "
1596                     "communications.");
1597             fr->pmePpCommGpu = std::make_unique<gmx::PmePpCommGpu>(
1598                     cr->mpi_comm_mysim,
1599                     cr->dd->pme_nodeid,
1600                     deviceStreamManager->context(),
1601                     deviceStreamManager->stream(DeviceStreamType::PmePpTransfer));
1602         }
1603
1604         fr->nbv = Nbnxm::init_nb_verlet(mdlog,
1605                                         inputrec.get(),
1606                                         fr,
1607                                         cr,
1608                                         *hwinfo_,
1609                                         runScheduleWork.simulationWork.useGpuNonbonded,
1610                                         deviceStreamManager.get(),
1611                                         &mtop,
1612                                         box,
1613                                         wcycle);
1614         // TODO: Move the logic below to a GPU bonded builder
1615         if (runScheduleWork.simulationWork.useGpuBonded)
1616         {
1617             GMX_RELEASE_ASSERT(deviceStreamManager != nullptr,
1618                                "GPU device stream manager should be valid in order to use GPU "
1619                                "version of bonded forces.");
1620             gpuBonded = std::make_unique<GpuBonded>(
1621                     mtop.ffparams,
1622                     fr->ic->epsfac * fr->fudgeQQ,
1623                     deviceStreamManager->context(),
1624                     deviceStreamManager->bondedStream(havePPDomainDecomposition(cr)),
1625                     wcycle);
1626             fr->gpuBonded = gpuBonded.get();
1627         }
1628
1629         /* Initialize the mdAtoms structure.
1630          * mdAtoms is not filled with atom data,
1631          * as this can not be done now with domain decomposition.
1632          */
1633         mdAtoms = makeMDAtoms(fplog, mtop, *inputrec, thisRankHasPmeGpuTask);
1634         if (globalState && thisRankHasPmeGpuTask)
1635         {
1636             // The pinning of coordinates in the global state object works, because we only use
1637             // PME on GPU without DD or on a separate PME rank, and because the local state pointer
1638             // points to the global state object without DD.
1639             // FIXME: MD and EM separately set up the local state - this should happen in the same
1640             // function, which should also perform the pinning.
1641             changePinningPolicy(&globalState->x, pme_get_pinning_policy());
1642         }
1643
1644         /* Initialize the virtual site communication */
1645         vsite = makeVirtualSitesHandler(mtop, cr, fr->pbcType);
1646
1647         calc_shifts(box, fr->shift_vec);
1648
1649         /* With periodic molecules the charge groups should be whole at start up
1650          * and the virtual sites should not be far from their proper positions.
1651          */
1652         if (!inputrec->bContinuation && MASTER(cr)
1653             && !(inputrec->pbcType != PbcType::No && inputrec->bPeriodicMols))
1654         {
1655             /* Make molecules whole at start of run */
1656             if (fr->pbcType != PbcType::No)
1657             {
1658                 do_pbc_first_mtop(fplog, inputrec->pbcType, box, &mtop, globalState->x.rvec_array());
1659             }
1660             if (vsite)
1661             {
1662                 /* Correct initial vsite positions are required
1663                  * for the initial distribution in the domain decomposition
1664                  * and for the initial shell prediction.
1665                  */
1666                 constructVirtualSitesGlobal(mtop, globalState->x);
1667             }
1668         }
1669
1670         if (EEL_PME(fr->ic->eeltype) || EVDW_PME(fr->ic->vdwtype))
1671         {
1672             ewaldcoeff_q  = fr->ic->ewaldcoeff_q;
1673             ewaldcoeff_lj = fr->ic->ewaldcoeff_lj;
1674         }
1675     }
1676     else
1677     {
1678         /* This is a PME only node */
1679
1680         GMX_ASSERT(globalState == nullptr,
1681                    "We don't need the state on a PME only rank and expect it to be unitialized");
1682
1683         ewaldcoeff_q  = calc_ewaldcoeff_q(inputrec->rcoulomb, inputrec->ewald_rtol);
1684         ewaldcoeff_lj = calc_ewaldcoeff_lj(inputrec->rvdw, inputrec->ewald_rtol_lj);
1685     }
1686
1687     gmx_pme_t* sepPmeData = nullptr;
1688     // This reference hides the fact that PME data is owned by runner on PME-only ranks and by forcerec on other ranks
1689     GMX_ASSERT(thisRankHasDuty(cr, DUTY_PP) == (fr != nullptr),
1690                "Double-checking that only PME-only ranks have no forcerec");
1691     gmx_pme_t*& pmedata = fr ? fr->pmedata : sepPmeData;
1692
1693     // TODO should live in ewald module once its testing is improved
1694     //
1695     // Later, this program could contain kernels that might be later
1696     // re-used as auto-tuning progresses, or subsequent simulations
1697     // are invoked.
1698     PmeGpuProgramStorage pmeGpuProgram;
1699     if (thisRankHasPmeGpuTask)
1700     {
1701         GMX_RELEASE_ASSERT(
1702                 (deviceStreamManager != nullptr),
1703                 "GPU device stream manager should be initialized in order to use GPU for PME.");
1704         GMX_RELEASE_ASSERT((deviceInfo != nullptr),
1705                            "GPU device should be initialized in order to use GPU for PME.");
1706         pmeGpuProgram = buildPmeGpuProgram(deviceStreamManager->context());
1707     }
1708
1709     /* Initiate PME if necessary,
1710      * either on all nodes or on dedicated PME nodes only. */
1711     if (EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype))
1712     {
1713         if (mdAtoms && mdAtoms->mdatoms())
1714         {
1715             nChargePerturbed = mdAtoms->mdatoms()->nChargePerturbed;
1716             if (EVDW_PME(inputrec->vdwtype))
1717             {
1718                 nTypePerturbed = mdAtoms->mdatoms()->nTypePerturbed;
1719             }
1720         }
1721         if (cr->npmenodes > 0)
1722         {
1723             /* The PME only nodes need to know nChargePerturbed(FEP on Q) and nTypePerturbed(FEP on LJ)*/
1724             gmx_bcast(sizeof(nChargePerturbed), &nChargePerturbed, cr->mpi_comm_mysim);
1725             gmx_bcast(sizeof(nTypePerturbed), &nTypePerturbed, cr->mpi_comm_mysim);
1726         }
1727
1728         if (thisRankHasDuty(cr, DUTY_PME))
1729         {
1730             try
1731             {
1732                 // TODO: This should be in the builder.
1733                 GMX_RELEASE_ASSERT(!runScheduleWork.simulationWork.useGpuPme
1734                                            || (deviceStreamManager != nullptr),
1735                                    "Device stream manager should be valid in order to use GPU "
1736                                    "version of PME.");
1737                 GMX_RELEASE_ASSERT(
1738                         !runScheduleWork.simulationWork.useGpuPme
1739                                 || deviceStreamManager->streamIsValid(DeviceStreamType::Pme),
1740                         "GPU PME stream should be valid in order to use GPU version of PME.");
1741
1742                 const DeviceContext* deviceContext = runScheduleWork.simulationWork.useGpuPme
1743                                                              ? &deviceStreamManager->context()
1744                                                              : nullptr;
1745                 const DeviceStream* pmeStream =
1746                         runScheduleWork.simulationWork.useGpuPme
1747                                 ? &deviceStreamManager->stream(DeviceStreamType::Pme)
1748                                 : nullptr;
1749
1750                 pmedata = gmx_pme_init(cr,
1751                                        getNumPmeDomains(cr->dd),
1752                                        inputrec.get(),
1753                                        nChargePerturbed != 0,
1754                                        nTypePerturbed != 0,
1755                                        mdrunOptions.reproducible,
1756                                        ewaldcoeff_q,
1757                                        ewaldcoeff_lj,
1758                                        gmx_omp_nthreads_get(emntPME),
1759                                        pmeRunMode,
1760                                        nullptr,
1761                                        deviceContext,
1762                                        pmeStream,
1763                                        pmeGpuProgram.get(),
1764                                        mdlog);
1765             }
1766             GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1767         }
1768     }
1769
1770
1771     if (EI_DYNAMICS(inputrec->eI))
1772     {
1773         /* Turn on signal handling on all nodes */
1774         /*
1775          * (A user signal from the PME nodes (if any)
1776          * is communicated to the PP nodes.
1777          */
1778         signal_handler_install();
1779     }
1780
1781     pull_t* pull_work = nullptr;
1782     if (thisRankHasDuty(cr, DUTY_PP))
1783     {
1784         /* Assumes uniform use of the number of OpenMP threads */
1785         walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(emntDefault));
1786
1787         if (inputrec->bPull)
1788         {
1789             /* Initialize pull code */
1790             pull_work = init_pull(fplog,
1791                                   inputrec->pull.get(),
1792                                   inputrec.get(),
1793                                   &mtop,
1794                                   cr,
1795                                   &atomSets,
1796                                   inputrec->fepvals->init_lambda);
1797             if (inputrec->pull->bXOutAverage || inputrec->pull->bFOutAverage)
1798             {
1799                 initPullHistory(pull_work, &observablesHistory);
1800             }
1801             if (EI_DYNAMICS(inputrec->eI) && MASTER(cr))
1802             {
1803                 init_pull_output_files(pull_work, filenames.size(), filenames.data(), oenv, startingBehavior);
1804             }
1805         }
1806
1807         std::unique_ptr<EnforcedRotation> enforcedRotation;
1808         if (inputrec->bRot)
1809         {
1810             /* Initialize enforced rotation code */
1811             enforcedRotation = init_rot(fplog,
1812                                         inputrec.get(),
1813                                         filenames.size(),
1814                                         filenames.data(),
1815                                         cr,
1816                                         &atomSets,
1817                                         globalState.get(),
1818                                         &mtop,
1819                                         oenv,
1820                                         mdrunOptions,
1821                                         startingBehavior);
1822         }
1823
1824         t_swap* swap = nullptr;
1825         if (inputrec->eSwapCoords != eswapNO)
1826         {
1827             /* Initialize ion swapping code */
1828             swap = init_swapcoords(fplog,
1829                                    inputrec.get(),
1830                                    opt2fn_master("-swap", filenames.size(), filenames.data(), cr),
1831                                    &mtop,
1832                                    globalState.get(),
1833                                    &observablesHistory,
1834                                    cr,
1835                                    &atomSets,
1836                                    oenv,
1837                                    mdrunOptions,
1838                                    startingBehavior);
1839         }
1840
1841         /* Let makeConstraints know whether we have essential dynamics constraints. */
1842         auto constr = makeConstraints(
1843                 mtop, *inputrec, pull_work, doEssentialDynamics, fplog, cr, ms, &nrnb, wcycle, fr->bMolPBC);
1844
1845         /* Energy terms and groups */
1846         gmx_enerdata_t enerd(mtop.groups.groups[SimulationAtomGroupType::EnergyOutput].size(),
1847                              inputrec->fepvals->n_lambda);
1848
1849         /* Kinetic energy data */
1850         gmx_ekindata_t ekind;
1851         init_ekindata(fplog, &(inputrec->opts), &ekind, inputrec->cos_accel);
1852
1853         /* Set up interactive MD (IMD) */
1854         auto imdSession = makeImdSession(inputrec.get(),
1855                                          cr,
1856                                          wcycle,
1857                                          &enerd,
1858                                          ms,
1859                                          &mtop,
1860                                          mdlog,
1861                                          MASTER(cr) ? globalState->x.rvec_array() : nullptr,
1862                                          filenames.size(),
1863                                          filenames.data(),
1864                                          oenv,
1865                                          mdrunOptions.imdOptions,
1866                                          startingBehavior);
1867
1868         if (DOMAINDECOMP(cr))
1869         {
1870             GMX_RELEASE_ASSERT(fr, "fr was NULL while cr->duty was DUTY_PP");
1871             /* This call is not included in init_domain_decomposition mainly
1872              * because fr->cginfo_mb is set later.
1873              */
1874             dd_init_bondeds(fplog,
1875                             cr->dd,
1876                             mtop,
1877                             vsite.get(),
1878                             inputrec.get(),
1879                             domdecOptions.checkBondedInteractions,
1880                             fr->cginfo_mb);
1881         }
1882
1883         if (runScheduleWork.simulationWork.useGpuBufferOps)
1884         {
1885             fr->gpuForceReduction[gmx::AtomLocality::Local] = std::make_unique<gmx::GpuForceReduction>(
1886                     deviceStreamManager->context(),
1887                     deviceStreamManager->stream(gmx::DeviceStreamType::NonBondedLocal),
1888                     wcycle);
1889             fr->gpuForceReduction[gmx::AtomLocality::NonLocal] = std::make_unique<gmx::GpuForceReduction>(
1890                     deviceStreamManager->context(),
1891                     deviceStreamManager->stream(gmx::DeviceStreamType::NonBondedNonLocal),
1892                     wcycle);
1893         }
1894
1895         std::unique_ptr<gmx::StatePropagatorDataGpu> stateGpu;
1896         if (gpusWereDetected
1897             && ((runScheduleWork.simulationWork.useGpuPme && thisRankHasDuty(cr, DUTY_PME))
1898                 || runScheduleWork.simulationWork.useGpuBufferOps))
1899         {
1900             GpuApiCallBehavior transferKind = (inputrec->eI == eiMD && !doRerun && !useModularSimulator)
1901                                                       ? GpuApiCallBehavior::Async
1902                                                       : GpuApiCallBehavior::Sync;
1903             GMX_RELEASE_ASSERT(deviceStreamManager != nullptr,
1904                                "GPU device stream manager should be initialized to use GPU.");
1905             stateGpu = std::make_unique<gmx::StatePropagatorDataGpu>(
1906                     *deviceStreamManager, transferKind, pme_gpu_get_block_size(fr->pmedata), wcycle);
1907             fr->stateGpu = stateGpu.get();
1908         }
1909
1910         GMX_ASSERT(stopHandlerBuilder_, "Runner must provide StopHandlerBuilder to simulator.");
1911         SimulatorBuilder simulatorBuilder;
1912
1913         simulatorBuilder.add(SimulatorStateData(globalState.get(), &observablesHistory, &enerd, &ekind));
1914         simulatorBuilder.add(std::move(membedHolder));
1915         simulatorBuilder.add(std::move(stopHandlerBuilder_));
1916         simulatorBuilder.add(SimulatorConfig(mdrunOptions, startingBehavior, &runScheduleWork));
1917
1918
1919         simulatorBuilder.add(SimulatorEnv(fplog, cr, ms, mdlog, oenv));
1920         simulatorBuilder.add(Profiling(&nrnb, walltime_accounting, wcycle));
1921         simulatorBuilder.add(ConstraintsParam(
1922                 constr.get(), enforcedRotation ? enforcedRotation->getLegacyEnfrot() : nullptr, vsite.get()));
1923         // TODO: Separate `fr` to a separate add, and make the `build` handle the coupling sensibly.
1924         simulatorBuilder.add(LegacyInput(
1925                 static_cast<int>(filenames.size()), filenames.data(), inputrec.get(), fr));
1926         simulatorBuilder.add(ReplicaExchangeParameters(replExParams));
1927         simulatorBuilder.add(InteractiveMD(imdSession.get()));
1928         simulatorBuilder.add(SimulatorModules(mdModules_->outputProvider(), mdModules_->notifier()));
1929         simulatorBuilder.add(CenterOfMassPulling(pull_work));
1930         // Todo move to an MDModule
1931         simulatorBuilder.add(IonSwapping(swap));
1932         simulatorBuilder.add(TopologyData(&mtop, mdAtoms.get()));
1933         simulatorBuilder.add(BoxDeformationHandle(deform.get()));
1934         simulatorBuilder.add(std::move(modularSimulatorCheckpointData));
1935
1936         // build and run simulator object based on user-input
1937         auto simulator = simulatorBuilder.build(useModularSimulator);
1938         simulator->run();
1939
1940         if (fr->pmePpCommGpu)
1941         {
1942             // destroy object since it is no longer required. (This needs to be done while the GPU context still exists.)
1943             fr->pmePpCommGpu.reset();
1944         }
1945
1946         if (inputrec->bPull)
1947         {
1948             finish_pull(pull_work);
1949         }
1950         finish_swapcoords(swap);
1951     }
1952     else
1953     {
1954         GMX_RELEASE_ASSERT(pmedata, "pmedata was NULL while cr->duty was not DUTY_PP");
1955         /* do PME only */
1956         walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(emntPME));
1957         gmx_pmeonly(pmedata,
1958                     cr,
1959                     &nrnb,
1960                     wcycle,
1961                     walltime_accounting,
1962                     inputrec.get(),
1963                     pmeRunMode,
1964                     deviceStreamManager.get());
1965     }
1966
1967     wallcycle_stop(wcycle, ewcRUN);
1968
1969     /* Finish up, write some stuff
1970      * if rerunMD, don't write last frame again
1971      */
1972     finish_run(fplog,
1973                mdlog,
1974                cr,
1975                inputrec.get(),
1976                &nrnb,
1977                wcycle,
1978                walltime_accounting,
1979                fr ? fr->nbv.get() : nullptr,
1980                pmedata,
1981                EI_DYNAMICS(inputrec->eI) && !isMultiSim(ms));
1982
1983     // clean up cycle counter
1984     wallcycle_destroy(wcycle);
1985
1986     deviceStreamManager.reset(nullptr);
1987     // Free PME data
1988     if (pmedata)
1989     {
1990         gmx_pme_destroy(pmedata);
1991         pmedata = nullptr;
1992     }
1993
1994     // FIXME: this is only here to manually unpin mdAtoms->chargeA_ and state->x,
1995     // before we destroy the GPU context(s)
1996     // Pinned buffers are associated with contexts in CUDA.
1997     // As soon as we destroy GPU contexts after mdrunner() exits, these lines should go.
1998     mdAtoms.reset(nullptr);
1999     globalState.reset(nullptr);
2000     mdModules_.reset(nullptr); // destruct force providers here as they might also use the GPU
2001     gpuBonded.reset(nullptr);
2002     /* Free pinned buffers in *fr */
2003     delete fr;
2004     fr = nullptr;
2005     // TODO convert to C++ so we can get rid of these frees
2006     sfree(disresdata);
2007     sfree(oriresdata);
2008
2009     if (!hwinfo_->deviceInfoList.empty())
2010     {
2011         /* stop the GPU profiler (only CUDA) */
2012         stopGpuProfiler();
2013     }
2014
2015     /* With tMPI we need to wait for all ranks to finish deallocation before
2016      * destroying the CUDA context as some tMPI ranks may be sharing
2017      * GPU and context.
2018      *
2019      * This is not a concern in OpenCL where we use one context per rank.
2020      *
2021      * Note: it is safe to not call the barrier on the ranks which do not use GPU,
2022      * but it is easier and more futureproof to call it on the whole node.
2023      *
2024      * Note that this function needs to be called even if GPUs are not used
2025      * in this run because the PME ranks have no knowledge of whether GPUs
2026      * are used or not, but all ranks need to enter the barrier below.
2027      * \todo Remove this physical node barrier after making sure
2028      * that it's not needed anymore (with a shared GPU run).
2029      */
2030     if (GMX_THREAD_MPI)
2031     {
2032         physicalNodeComm.barrier();
2033     }
2034     releaseDevice(deviceInfo);
2035
2036     /* Does what it says */
2037     print_date_and_time(fplog, cr->nodeid, "Finished mdrun", gmx_gettime());
2038     walltime_accounting_destroy(walltime_accounting);
2039
2040     // Ensure log file content is written
2041     if (logFileHandle)
2042     {
2043         gmx_fio_flush(logFileHandle);
2044     }
2045
2046     /* Reset FPEs (important for unit tests) by disabling them. Assumes no
2047      * exceptions were enabled before function was called. */
2048     if (bEnableFPE)
2049     {
2050         gmx_fedisableexcept();
2051     }
2052
2053     auto rc = static_cast<int>(gmx_get_stop_condition());
2054
2055 #if GMX_THREAD_MPI
2056     /* we need to join all threads. The sub-threads join when they
2057        exit this function, but the master thread needs to be told to
2058        wait for that. */
2059     if (MASTER(cr))
2060     {
2061         tMPI_Finalize();
2062     }
2063 #endif
2064     return rc;
2065 } // namespace gmx
2066
2067 Mdrunner::~Mdrunner()
2068 {
2069     // Clean up of the Manager.
2070     // This will end up getting called on every thread-MPI rank, which is unnecessary,
2071     // but okay as long as threads synchronize some time before adding or accessing
2072     // a new set of restraints.
2073     if (restraintManager_)
2074     {
2075         restraintManager_->clear();
2076         GMX_ASSERT(restraintManager_->countRestraints() == 0,
2077                    "restraints added during runner life time should be cleared at runner "
2078                    "destruction.");
2079     }
2080 };
2081
2082 void Mdrunner::addPotential(std::shared_ptr<gmx::IRestraintPotential> puller, const std::string& name)
2083 {
2084     GMX_ASSERT(restraintManager_, "Mdrunner must have a restraint manager.");
2085     // Not sure if this should be logged through the md logger or something else,
2086     // but it is helpful to have some sort of INFO level message sent somewhere.
2087     //    std::cout << "Registering restraint named " << name << std::endl;
2088
2089     // When multiple restraints are used, it may be wasteful to register them separately.
2090     // Maybe instead register an entire Restraint Manager as a force provider.
2091     restraintManager_->addToSpec(std::move(puller), name);
2092 }
2093
2094 Mdrunner::Mdrunner(std::unique_ptr<MDModules> mdModules) : mdModules_(std::move(mdModules)) {}
2095
2096 Mdrunner::Mdrunner(Mdrunner&&) noexcept = default;
2097
2098 //NOLINTNEXTLINE(performance-noexcept-move-constructor) working around GCC bug 58265 in CentOS 7
2099 Mdrunner& Mdrunner::operator=(Mdrunner&& /*handle*/) noexcept(BUGFREE_NOEXCEPT_STRING) = default;
2100
2101 class Mdrunner::BuilderImplementation
2102 {
2103 public:
2104     BuilderImplementation() = delete;
2105     BuilderImplementation(std::unique_ptr<MDModules> mdModules, compat::not_null<SimulationContext*> context);
2106     ~BuilderImplementation();
2107
2108     BuilderImplementation& setExtraMdrunOptions(const MdrunOptions& options,
2109                                                 real                forceWarningThreshold,
2110                                                 StartingBehavior    startingBehavior);
2111
2112     void addHardwareDetectionResult(const gmx_hw_info_t* hwinfo);
2113
2114     void addDomdec(const DomdecOptions& options);
2115
2116     void addInput(SimulationInputHandle inputHolder);
2117
2118     void addVerletList(int nstlist);
2119
2120     void addReplicaExchange(const ReplicaExchangeParameters& params);
2121
2122     void addNonBonded(const char* nbpu_opt);
2123
2124     void addPME(const char* pme_opt_, const char* pme_fft_opt_);
2125
2126     void addBondedTaskAssignment(const char* bonded_opt);
2127
2128     void addUpdateTaskAssignment(const char* update_opt);
2129
2130     void addHardwareOptions(const gmx_hw_opt_t& hardwareOptions);
2131
2132     void addFilenames(ArrayRef<const t_filenm> filenames);
2133
2134     void addOutputEnvironment(gmx_output_env_t* outputEnvironment);
2135
2136     void addLogFile(t_fileio* logFileHandle);
2137
2138     void addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder);
2139
2140     Mdrunner build();
2141
2142 private:
2143     // Default parameters copied from runner.h
2144     // \todo Clarify source(s) of default parameters.
2145
2146     const char* nbpu_opt_    = nullptr;
2147     const char* pme_opt_     = nullptr;
2148     const char* pme_fft_opt_ = nullptr;
2149     const char* bonded_opt_  = nullptr;
2150     const char* update_opt_  = nullptr;
2151
2152     MdrunOptions mdrunOptions_;
2153
2154     DomdecOptions domdecOptions_;
2155
2156     ReplicaExchangeParameters replicaExchangeParameters_;
2157
2158     //! Command-line override for the duration of a neighbor list with the Verlet scheme.
2159     int nstlist_ = 0;
2160
2161     //! World communicator, used for hardware detection and task assignment
2162     MPI_Comm libraryWorldCommunicator_ = MPI_COMM_NULL;
2163
2164     //! Multisim communicator handle.
2165     gmx_multisim_t* multiSimulation_;
2166
2167     //! mdrun communicator
2168     MPI_Comm simulationCommunicator_ = MPI_COMM_NULL;
2169
2170     //! Print a warning if any force is larger than this (in kJ/mol nm).
2171     real forceWarningThreshold_ = -1;
2172
2173     //! Whether the simulation will start afresh, or restart with/without appending.
2174     StartingBehavior startingBehavior_ = StartingBehavior::NewSimulation;
2175
2176     //! The modules that comprise the functionality of mdrun.
2177     std::unique_ptr<MDModules> mdModules_;
2178
2179     //! Detected hardware.
2180     const gmx_hw_info_t* hwinfo_ = nullptr;
2181
2182     //! \brief Parallelism information.
2183     gmx_hw_opt_t hardwareOptions_;
2184
2185     //! filename options for simulation.
2186     ArrayRef<const t_filenm> filenames_;
2187
2188     /*! \brief Handle to output environment.
2189      *
2190      * \todo gmx_output_env_t needs lifetime management.
2191      */
2192     gmx_output_env_t* outputEnvironment_ = nullptr;
2193
2194     /*! \brief Non-owning handle to MD log file.
2195      *
2196      * \todo Context should own output facilities for client.
2197      * \todo Improve log file handle management.
2198      * \internal
2199      * Code managing the FILE* relies on the ability to set it to
2200      * nullptr to check whether the filehandle is valid.
2201      */
2202     t_fileio* logFileHandle_ = nullptr;
2203
2204     /*!
2205      * \brief Builder for simulation stop signal handler.
2206      */
2207     std::unique_ptr<StopHandlerBuilder> stopHandlerBuilder_ = nullptr;
2208
2209     /*!
2210      * \brief Sources for initial simulation state.
2211      *
2212      * See issue #3652 for near-term refinements to the SimulationInput interface.
2213      *
2214      * See issue #3379 for broader discussion on API aspects of simulation inputs and outputs.
2215      */
2216     SimulationInputHandle inputHolder_;
2217 };
2218
2219 Mdrunner::BuilderImplementation::BuilderImplementation(std::unique_ptr<MDModules> mdModules,
2220                                                        compat::not_null<SimulationContext*> context) :
2221     mdModules_(std::move(mdModules))
2222 {
2223     libraryWorldCommunicator_ = context->libraryWorldCommunicator_;
2224     simulationCommunicator_   = context->simulationCommunicator_;
2225     multiSimulation_          = context->multiSimulation_.get();
2226 }
2227
2228 Mdrunner::BuilderImplementation::~BuilderImplementation() = default;
2229
2230 Mdrunner::BuilderImplementation&
2231 Mdrunner::BuilderImplementation::setExtraMdrunOptions(const MdrunOptions&    options,
2232                                                       const real             forceWarningThreshold,
2233                                                       const StartingBehavior startingBehavior)
2234 {
2235     mdrunOptions_          = options;
2236     forceWarningThreshold_ = forceWarningThreshold;
2237     startingBehavior_      = startingBehavior;
2238     return *this;
2239 }
2240
2241 void Mdrunner::BuilderImplementation::addDomdec(const DomdecOptions& options)
2242 {
2243     domdecOptions_ = options;
2244 }
2245
2246 void Mdrunner::BuilderImplementation::addVerletList(int nstlist)
2247 {
2248     nstlist_ = nstlist;
2249 }
2250
2251 void Mdrunner::BuilderImplementation::addReplicaExchange(const ReplicaExchangeParameters& params)
2252 {
2253     replicaExchangeParameters_ = params;
2254 }
2255
2256 Mdrunner Mdrunner::BuilderImplementation::build()
2257 {
2258     auto newRunner = Mdrunner(std::move(mdModules_));
2259
2260     newRunner.mdrunOptions     = mdrunOptions_;
2261     newRunner.pforce           = forceWarningThreshold_;
2262     newRunner.startingBehavior = startingBehavior_;
2263     newRunner.domdecOptions    = domdecOptions_;
2264
2265     // \todo determine an invariant to check or confirm that all gmx_hw_opt_t objects are valid
2266     newRunner.hw_opt = hardwareOptions_;
2267
2268     // No invariant to check. This parameter exists to optionally override other behavior.
2269     newRunner.nstlist_cmdline = nstlist_;
2270
2271     newRunner.replExParams = replicaExchangeParameters_;
2272
2273     newRunner.filenames = filenames_;
2274
2275     newRunner.libraryWorldCommunicator = libraryWorldCommunicator_;
2276
2277     newRunner.simulationCommunicator = simulationCommunicator_;
2278
2279     // nullptr is a valid value for the multisim handle
2280     newRunner.ms = multiSimulation_;
2281
2282     if (hwinfo_)
2283     {
2284         newRunner.hwinfo_ = hwinfo_;
2285     }
2286     else
2287     {
2288         GMX_THROW(gmx::APIError(
2289                 "MdrunnerBuilder::addHardwareDetectionResult() is required before build()"));
2290     }
2291
2292     if (inputHolder_)
2293     {
2294         newRunner.inputHolder_ = std::move(inputHolder_);
2295     }
2296     else
2297     {
2298         GMX_THROW(gmx::APIError("MdrunnerBuilder::addInput() is required before build()."));
2299     }
2300
2301     // \todo Clarify ownership and lifetime management for gmx_output_env_t
2302     // \todo Update sanity checking when output environment has clearly specified invariants.
2303     // Initialization and default values for oenv are not well specified in the current version.
2304     if (outputEnvironment_)
2305     {
2306         newRunner.oenv = outputEnvironment_;
2307     }
2308     else
2309     {
2310         GMX_THROW(gmx::APIError(
2311                 "MdrunnerBuilder::addOutputEnvironment() is required before build()"));
2312     }
2313
2314     newRunner.logFileHandle = logFileHandle_;
2315
2316     if (nbpu_opt_)
2317     {
2318         newRunner.nbpu_opt = nbpu_opt_;
2319     }
2320     else
2321     {
2322         GMX_THROW(gmx::APIError("MdrunnerBuilder::addNonBonded() is required before build()"));
2323     }
2324
2325     if (pme_opt_ && pme_fft_opt_)
2326     {
2327         newRunner.pme_opt     = pme_opt_;
2328         newRunner.pme_fft_opt = pme_fft_opt_;
2329     }
2330     else
2331     {
2332         GMX_THROW(gmx::APIError("MdrunnerBuilder::addElectrostatics() is required before build()"));
2333     }
2334
2335     if (bonded_opt_)
2336     {
2337         newRunner.bonded_opt = bonded_opt_;
2338     }
2339     else
2340     {
2341         GMX_THROW(gmx::APIError(
2342                 "MdrunnerBuilder::addBondedTaskAssignment() is required before build()"));
2343     }
2344
2345     if (update_opt_)
2346     {
2347         newRunner.update_opt = update_opt_;
2348     }
2349     else
2350     {
2351         GMX_THROW(gmx::APIError(
2352                 "MdrunnerBuilder::addUpdateTaskAssignment() is required before build()  "));
2353     }
2354
2355
2356     newRunner.restraintManager_ = std::make_unique<gmx::RestraintManager>();
2357
2358     if (stopHandlerBuilder_)
2359     {
2360         newRunner.stopHandlerBuilder_ = std::move(stopHandlerBuilder_);
2361     }
2362     else
2363     {
2364         newRunner.stopHandlerBuilder_ = std::make_unique<StopHandlerBuilder>();
2365     }
2366
2367     return newRunner;
2368 }
2369
2370 void Mdrunner::BuilderImplementation::addHardwareDetectionResult(const gmx_hw_info_t* hwinfo)
2371 {
2372     hwinfo_ = hwinfo;
2373 }
2374
2375 void Mdrunner::BuilderImplementation::addNonBonded(const char* nbpu_opt)
2376 {
2377     nbpu_opt_ = nbpu_opt;
2378 }
2379
2380 void Mdrunner::BuilderImplementation::addPME(const char* pme_opt, const char* pme_fft_opt)
2381 {
2382     pme_opt_     = pme_opt;
2383     pme_fft_opt_ = pme_fft_opt;
2384 }
2385
2386 void Mdrunner::BuilderImplementation::addBondedTaskAssignment(const char* bonded_opt)
2387 {
2388     bonded_opt_ = bonded_opt;
2389 }
2390
2391 void Mdrunner::BuilderImplementation::addUpdateTaskAssignment(const char* update_opt)
2392 {
2393     update_opt_ = update_opt;
2394 }
2395
2396 void Mdrunner::BuilderImplementation::addHardwareOptions(const gmx_hw_opt_t& hardwareOptions)
2397 {
2398     hardwareOptions_ = hardwareOptions;
2399 }
2400
2401 void Mdrunner::BuilderImplementation::addFilenames(ArrayRef<const t_filenm> filenames)
2402 {
2403     filenames_ = filenames;
2404 }
2405
2406 void Mdrunner::BuilderImplementation::addOutputEnvironment(gmx_output_env_t* outputEnvironment)
2407 {
2408     outputEnvironment_ = outputEnvironment;
2409 }
2410
2411 void Mdrunner::BuilderImplementation::addLogFile(t_fileio* logFileHandle)
2412 {
2413     logFileHandle_ = logFileHandle;
2414 }
2415
2416 void Mdrunner::BuilderImplementation::addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder)
2417 {
2418     stopHandlerBuilder_ = std::move(builder);
2419 }
2420
2421 void Mdrunner::BuilderImplementation::addInput(SimulationInputHandle inputHolder)
2422 {
2423     inputHolder_ = std::move(inputHolder);
2424 }
2425
2426 MdrunnerBuilder::MdrunnerBuilder(std::unique_ptr<MDModules>           mdModules,
2427                                  compat::not_null<SimulationContext*> context) :
2428     impl_{ std::make_unique<Mdrunner::BuilderImplementation>(std::move(mdModules), context) }
2429 {
2430 }
2431
2432 MdrunnerBuilder::~MdrunnerBuilder() = default;
2433
2434 MdrunnerBuilder& MdrunnerBuilder::addHardwareDetectionResult(const gmx_hw_info_t* hwinfo)
2435 {
2436     impl_->addHardwareDetectionResult(hwinfo);
2437     return *this;
2438 }
2439
2440 MdrunnerBuilder& MdrunnerBuilder::addSimulationMethod(const MdrunOptions&    options,
2441                                                       real                   forceWarningThreshold,
2442                                                       const StartingBehavior startingBehavior)
2443 {
2444     impl_->setExtraMdrunOptions(options, forceWarningThreshold, startingBehavior);
2445     return *this;
2446 }
2447
2448 MdrunnerBuilder& MdrunnerBuilder::addDomainDecomposition(const DomdecOptions& options)
2449 {
2450     impl_->addDomdec(options);
2451     return *this;
2452 }
2453
2454 MdrunnerBuilder& MdrunnerBuilder::addNeighborList(int nstlist)
2455 {
2456     impl_->addVerletList(nstlist);
2457     return *this;
2458 }
2459
2460 MdrunnerBuilder& MdrunnerBuilder::addReplicaExchange(const ReplicaExchangeParameters& params)
2461 {
2462     impl_->addReplicaExchange(params);
2463     return *this;
2464 }
2465
2466 MdrunnerBuilder& MdrunnerBuilder::addNonBonded(const char* nbpu_opt)
2467 {
2468     impl_->addNonBonded(nbpu_opt);
2469     return *this;
2470 }
2471
2472 MdrunnerBuilder& MdrunnerBuilder::addElectrostatics(const char* pme_opt, const char* pme_fft_opt)
2473 {
2474     // The builder method may become more general in the future, but in this version,
2475     // parameters for PME electrostatics are both required and the only parameters
2476     // available.
2477     if (pme_opt && pme_fft_opt)
2478     {
2479         impl_->addPME(pme_opt, pme_fft_opt);
2480     }
2481     else
2482     {
2483         GMX_THROW(
2484                 gmx::InvalidInputError("addElectrostatics() arguments must be non-null pointers."));
2485     }
2486     return *this;
2487 }
2488
2489 MdrunnerBuilder& MdrunnerBuilder::addBondedTaskAssignment(const char* bonded_opt)
2490 {
2491     impl_->addBondedTaskAssignment(bonded_opt);
2492     return *this;
2493 }
2494
2495 MdrunnerBuilder& MdrunnerBuilder::addUpdateTaskAssignment(const char* update_opt)
2496 {
2497     impl_->addUpdateTaskAssignment(update_opt);
2498     return *this;
2499 }
2500
2501 Mdrunner MdrunnerBuilder::build()
2502 {
2503     return impl_->build();
2504 }
2505
2506 MdrunnerBuilder& MdrunnerBuilder::addHardwareOptions(const gmx_hw_opt_t& hardwareOptions)
2507 {
2508     impl_->addHardwareOptions(hardwareOptions);
2509     return *this;
2510 }
2511
2512 MdrunnerBuilder& MdrunnerBuilder::addFilenames(ArrayRef<const t_filenm> filenames)
2513 {
2514     impl_->addFilenames(filenames);
2515     return *this;
2516 }
2517
2518 MdrunnerBuilder& MdrunnerBuilder::addOutputEnvironment(gmx_output_env_t* outputEnvironment)
2519 {
2520     impl_->addOutputEnvironment(outputEnvironment);
2521     return *this;
2522 }
2523
2524 MdrunnerBuilder& MdrunnerBuilder::addLogFile(t_fileio* logFileHandle)
2525 {
2526     impl_->addLogFile(logFileHandle);
2527     return *this;
2528 }
2529
2530 MdrunnerBuilder& MdrunnerBuilder::addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder)
2531 {
2532     impl_->addStopHandlerBuilder(std::move(builder));
2533     return *this;
2534 }
2535
2536 MdrunnerBuilder& MdrunnerBuilder::addInput(SimulationInputHandle input)
2537 {
2538     impl_->addInput(std::move(input));
2539     return *this;
2540 }
2541
2542 MdrunnerBuilder::MdrunnerBuilder(MdrunnerBuilder&&) noexcept = default;
2543
2544 MdrunnerBuilder& MdrunnerBuilder::operator=(MdrunnerBuilder&&) noexcept = default;
2545
2546 } // namespace gmx