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) 2011,2012,2013,2014,2015,2016,2017,2018,2019, by the GROMACS development team, led by
   7  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   8  * and including many others, as listed in the AUTHORS file in the
   9  * top-level source directory and at http://www.gromacs.org.
  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/domdec.h"
  61 #include "gromacs/domdec/domdec_struct.h"
  62 #include "gromacs/domdec/gpuhaloexchange.h"
  63 #include "gromacs/domdec/localatomsetmanager.h"
  64 #include "gromacs/domdec/partition.h"
  65 #include "gromacs/ewald/ewald_utils.h"
  66 #include "gromacs/ewald/pme.h"
  67 #include "gromacs/ewald/pme_gpu_program.h"
  68 #include "gromacs/fileio/checkpoint.h"
  69 #include "gromacs/fileio/gmxfio.h"
  70 #include "gromacs/fileio/oenv.h"
  71 #include "gromacs/fileio/tpxio.h"
  72 #include "gromacs/gmxlib/network.h"
  73 #include "gromacs/gmxlib/nrnb.h"
  74 #include "gromacs/gpu_utils/gpu_utils.h"
  75 #include "gromacs/hardware/cpuinfo.h"
  76 #include "gromacs/hardware/detecthardware.h"
  77 #include "gromacs/hardware/printhardware.h"
  78 #include "gromacs/imd/imd.h"
  79 #include "gromacs/listed_forces/disre.h"
  80 #include "gromacs/listed_forces/gpubonded.h"
  81 #include "gromacs/listed_forces/orires.h"
  82 #include "gromacs/math/functions.h"
  83 #include "gromacs/math/utilities.h"
  84 #include "gromacs/math/vec.h"
  85 #include "gromacs/mdlib/boxdeformation.h"
  86 #include "gromacs/mdlib/broadcaststructs.h"
  87 #include "gromacs/mdlib/calc_verletbuf.h"
  88 #include "gromacs/mdlib/dispersioncorrection.h"
  89 #include "gromacs/mdlib/enerdata_utils.h"
  90 #include "gromacs/mdlib/force.h"
  91 #include "gromacs/mdlib/forcerec.h"
  92 #include "gromacs/mdlib/gmx_omp_nthreads.h"
  93 #include "gromacs/mdlib/makeconstraints.h"
  94 #include "gromacs/mdlib/md_support.h"
  95 #include "gromacs/mdlib/mdatoms.h"
  96 #include "gromacs/mdlib/membed.h"
  97 #include "gromacs/mdlib/qmmm.h"
  98 #include "gromacs/mdlib/sighandler.h"
  99 #include "gromacs/mdlib/stophandler.h"
 100 #include "gromacs/mdrun/mdmodules.h"
 101 #include "gromacs/mdrun/simulationcontext.h"
 102 #include "gromacs/mdrunutility/handlerestart.h"
 103 #include "gromacs/mdrunutility/logging.h"
 104 #include "gromacs/mdrunutility/multisim.h"
 105 #include "gromacs/mdrunutility/printtime.h"
 106 #include "gromacs/mdrunutility/threadaffinity.h"
 107 #include "gromacs/mdtypes/commrec.h"
 108 #include "gromacs/mdtypes/enerdata.h"
 109 #include "gromacs/mdtypes/fcdata.h"
 110 #include "gromacs/mdtypes/group.h"
 111 #include "gromacs/mdtypes/inputrec.h"
 112 #include "gromacs/mdtypes/md_enums.h"
 113 #include "gromacs/mdtypes/mdrunoptions.h"
 114 #include "gromacs/mdtypes/observableshistory.h"
 115 #include "gromacs/mdtypes/simulation_workload.h"
 116 #include "gromacs/mdtypes/state.h"
 117 #include "gromacs/nbnxm/gpu_data_mgmt.h"
 118 #include "gromacs/nbnxm/nbnxm.h"
 119 #include "gromacs/nbnxm/pairlist_tuning.h"
 120 #include "gromacs/pbcutil/pbc.h"
 121 #include "gromacs/pulling/output.h"
 122 #include "gromacs/pulling/pull.h"
 123 #include "gromacs/pulling/pull_rotation.h"
 124 #include "gromacs/restraint/manager.h"
 125 #include "gromacs/restraint/restraintmdmodule.h"
 126 #include "gromacs/restraint/restraintpotential.h"
 127 #include "gromacs/swap/swapcoords.h"
 128 #include "gromacs/taskassignment/decidegpuusage.h"
 129 #include "gromacs/taskassignment/resourcedivision.h"
 130 #include "gromacs/taskassignment/taskassignment.h"
 131 #include "gromacs/taskassignment/usergpuids.h"
 132 #include "gromacs/timing/gpu_timing.h"
 133 #include "gromacs/timing/wallcycle.h"
 134 #include "gromacs/timing/wallcyclereporting.h"
 135 #include "gromacs/topology/mtop_util.h"
 136 #include "gromacs/trajectory/trajectoryframe.h"
 137 #include "gromacs/utility/basenetwork.h"
 138 #include "gromacs/utility/cstringutil.h"
 139 #include "gromacs/utility/exceptions.h"
 140 #include "gromacs/utility/fatalerror.h"
 141 #include "gromacs/utility/filestream.h"
 142 #include "gromacs/utility/gmxassert.h"
 143 #include "gromacs/utility/gmxmpi.h"
 144 #include "gromacs/utility/keyvaluetree.h"
 145 #include "gromacs/utility/logger.h"
 146 #include "gromacs/utility/loggerbuilder.h"
 147 #include "gromacs/utility/mdmodulenotification.h"
 148 #include "gromacs/utility/physicalnodecommunicator.h"
 149 #include "gromacs/utility/pleasecite.h"
 150 #include "gromacs/utility/programcontext.h"
 151 #include "gromacs/utility/smalloc.h"
 152 #include "gromacs/utility/stringutil.h"
 153
 154 #include "isimulator.h"
 155 #include "replicaexchange.h"
 156 #include "simulatorbuilder.h"
 157
 158 #if GMX_FAHCORE
 159 #include "corewrap.h"
 160 #endif
 161
 162 namespace gmx
 163 {
 164
 165 /*! \brief environment variable to enable GPU P2P communication */
 166 static const bool c_enableGpuHaloExchange = (getenv("GMX_GPU_DD_COMMS") != nullptr)
 167     && GMX_THREAD_MPI && (GMX_GPU == GMX_GPU_CUDA);
 168
 169 /*! \brief Manage any development feature flag variables encountered
 170  *
 171  * The use of dev features indicated by environment variables is
 172  * logged in order to ensure that runs with such featrues enabled can
 173  * be identified from their log and standard output. Any cross
 174  * dependencies are also checked, and if unsatisified, a fatal error
 175  * issued.
 176  *
 177  * \param[in]  mdlog        Logger object.
 178  */
 179 static void manageDevelopmentFeatures(const gmx::MDLogger &mdlog)
 180 {
 181     const bool enableGpuBufOps       = (getenv("GMX_USE_GPU_BUFFER_OPS") != nullptr);
 182     const bool useGpuUpdateConstrain = (getenv("GMX_UPDATE_CONSTRAIN_GPU") != nullptr);
 183     const bool enableGpuHaloExchange = (getenv("GMX_GPU_DD_COMMS") != nullptr && GMX_THREAD_MPI && (GMX_GPU == GMX_GPU_CUDA));
 184
 185     if (enableGpuBufOps)
 186     {
 187         GMX_LOG(mdlog.warning).asParagraph().appendTextFormatted(
 188                 "NOTE: This run uses the 'GPU buffer ops' feature, enabled by the GMX_USE_GPU_BUFFER_OPS environment variable.");
 189     }
 190
 191     if (enableGpuHaloExchange)
 192     {
 193         if (!enableGpuBufOps)
 194         {
 195             gmx_fatal(FARGS, "Cannot enable GPU halo exchange without GPU buffer operations, set GMX_USE_GPU_BUFFER_OPS=1\n");
 196         }
 197         GMX_LOG(mdlog.warning).asParagraph().appendTextFormatted(
 198                 "NOTE: This run uses the 'GPU halo exchange' feature, enabled by the GMX_GPU_DD_COMMS environment variable.");
 199     }
 200
 201     if (useGpuUpdateConstrain)
 202     {
 203         GMX_LOG(mdlog.warning).asParagraph().appendTextFormatted(
 204                 "NOTE: This run uses the 'GPU update/constraints' feature, enabled by the GMX_UPDATE_CONSTRAIN_GPU environment variable.");
 205     }
 206
 207 }
 208
 209 /*! \brief Barrier for safe simultaneous thread access to mdrunner data
 210  *
 211  * Used to ensure that the master thread does not modify mdrunner during copy
 212  * on the spawned threads. */
 213 static void threadMpiMdrunnerAccessBarrier()
 214 {
 215 #if GMX_THREAD_MPI
 216     MPI_Barrier(MPI_COMM_WORLD);
 217 #endif
 218 }
 219
 220 Mdrunner Mdrunner::cloneOnSpawnedThread() const
 221 {
 222     auto newRunner = Mdrunner(std::make_unique<MDModules>());
 223
 224     // All runners in the same process share a restraint manager resource because it is
 225     // part of the interface to the client code, which is associated only with the
 226     // original thread. Handles to the same resources can be obtained by copy.
 227     {
 228         newRunner.restraintManager_ = std::make_unique<RestraintManager>(*restraintManager_);
 229     }
 230
 231     // Copy original cr pointer before master thread can pass the thread barrier
 232     newRunner.cr  = reinitialize_commrec_for_this_thread(cr);
 233
 234     // Copy members of master runner.
 235     // \todo Replace with builder when Simulation context and/or runner phases are better defined.
 236     // Ref https://redmine.gromacs.org/issues/2587 and https://redmine.gromacs.org/issues/2375
 237     newRunner.hw_opt    = hw_opt;
 238     newRunner.filenames = filenames;
 239
 240     newRunner.oenv                = oenv;
 241     newRunner.mdrunOptions        = mdrunOptions;
 242     newRunner.domdecOptions       = domdecOptions;
 243     newRunner.nbpu_opt            = nbpu_opt;
 244     newRunner.pme_opt             = pme_opt;
 245     newRunner.pme_fft_opt         = pme_fft_opt;
 246     newRunner.bonded_opt          = bonded_opt;
 247     newRunner.update_opt          = update_opt;
 248     newRunner.nstlist_cmdline     = nstlist_cmdline;
 249     newRunner.replExParams        = replExParams;
 250     newRunner.pforce              = pforce;
 251     newRunner.ms                  = ms;
 252     newRunner.startingBehavior    = startingBehavior;
 253     newRunner.stopHandlerBuilder_ = std::make_unique<StopHandlerBuilder>(*stopHandlerBuilder_);
 254
 255     threadMpiMdrunnerAccessBarrier();
 256
 257     GMX_RELEASE_ASSERT(!MASTER(newRunner.cr), "cloneOnSpawnedThread should only be called on spawned threads");
 258
 259     return newRunner;
 260 }
 261
 262 /*! \brief The callback used for running on spawned threads.
 263  *
 264  * Obtains the pointer to the master mdrunner object from the one
 265  * argument permitted to the thread-launch API call, copies it to make
 266  * a new runner for this thread, reinitializes necessary data, and
 267  * proceeds to the simulation. */
 268 static void mdrunner_start_fn(const void *arg)
 269 {
 270     try
 271     {
 272         auto masterMdrunner = reinterpret_cast<const gmx::Mdrunner *>(arg);
 273         /* copy the arg list to make sure that it's thread-local. This
 274            doesn't copy pointed-to items, of course; fnm, cr and fplog
 275            are reset in the call below, all others should be const. */
 276         gmx::Mdrunner mdrunner = masterMdrunner->cloneOnSpawnedThread();
 277         mdrunner.mdrunner();
 278     }
 279     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
 280 }
 281
 282
 283 /*! \brief Start thread-MPI threads.
 284  *
 285  * Called by mdrunner() to start a specific number of threads
 286  * (including the main thread) for thread-parallel runs. This in turn
 287  * calls mdrunner() for each thread. All options are the same as for
 288  * mdrunner(). */
 289 t_commrec *Mdrunner::spawnThreads(int numThreadsToLaunch) const
 290 {
 291
 292     /* first check whether we even need to start tMPI */
 293     if (numThreadsToLaunch < 2)
 294     {
 295         return cr;
 296     }
 297
 298 #if GMX_THREAD_MPI
 299     /* now spawn new threads that start mdrunner_start_fn(), while
 300        the main thread returns. Thread affinity is handled later. */
 301     if (tMPI_Init_fn(TRUE, numThreadsToLaunch, TMPI_AFFINITY_NONE,
 302                      mdrunner_start_fn, static_cast<const void*>(this)) != TMPI_SUCCESS)
 303     {
 304         GMX_THROW(gmx::InternalError("Failed to spawn thread-MPI threads"));
 305     }
 306
 307     threadMpiMdrunnerAccessBarrier();
 308 #else
 309     GMX_UNUSED_VALUE(mdrunner_start_fn);
 310 #endif
 311
 312     return reinitialize_commrec_for_this_thread(cr);
 313 }
 314
 315 }  // namespace gmx
 316
 317 /*! \brief Initialize variables for Verlet scheme simulation */
 318 static void prepare_verlet_scheme(FILE                           *fplog,
 319                                   t_commrec                      *cr,
 320                                   t_inputrec                     *ir,
 321                                   int                             nstlist_cmdline,
 322                                   const gmx_mtop_t               *mtop,
 323                                   const matrix                    box,
 324                                   bool                            makeGpuPairList,
 325                                   const gmx::CpuInfo             &cpuinfo)
 326 {
 327     /* For NVE simulations, we will retain the initial list buffer */
 328     if (EI_DYNAMICS(ir->eI) &&
 329         ir->verletbuf_tol > 0 &&
 330         !(EI_MD(ir->eI) && ir->etc == etcNO))
 331     {
 332         /* Update the Verlet buffer size for the current run setup */
 333
 334         /* Here we assume SIMD-enabled kernels are being used. But as currently
 335          * calc_verlet_buffer_size gives the same results for 4x8 and 4x4
 336          * and 4x2 gives a larger buffer than 4x4, this is ok.
 337          */
 338         ListSetupType      listType  = (makeGpuPairList ? ListSetupType::Gpu : ListSetupType::CpuSimdWhenSupported);
 339         VerletbufListSetup listSetup = verletbufGetSafeListSetup(listType);
 340
 341         const real         rlist_new =
 342             calcVerletBufferSize(*mtop, det(box), *ir, ir->nstlist, ir->nstlist - 1, -1, listSetup);
 343
 344         if (rlist_new != ir->rlist)
 345         {
 346             if (fplog != nullptr)
 347             {
 348                 fprintf(fplog, "\nChanging rlist from %g to %g for non-bonded %dx%d atom kernels\n\n",
 349                         ir->rlist, rlist_new,
 350                         listSetup.cluster_size_i, listSetup.cluster_size_j);
 351             }
 352             ir->rlist     = rlist_new;
 353         }
 354     }
 355
 356     if (nstlist_cmdline > 0 && (!EI_DYNAMICS(ir->eI) || ir->verletbuf_tol <= 0))
 357     {
 358         gmx_fatal(FARGS, "Can not set nstlist without %s",
 359                   !EI_DYNAMICS(ir->eI) ? "dynamics" : "verlet-buffer-tolerance");
 360     }
 361
 362     if (EI_DYNAMICS(ir->eI))
 363     {
 364         /* Set or try nstlist values */
 365         increaseNstlist(fplog, cr, ir, nstlist_cmdline, mtop, box, makeGpuPairList, cpuinfo);
 366     }
 367 }
 368
 369 /*! \brief Override the nslist value in inputrec
 370  *
 371  * with value passed on the command line (if any)
 372  */
 373 static void override_nsteps_cmdline(const gmx::MDLogger &mdlog,
 374                                     int64_t              nsteps_cmdline,
 375                                     t_inputrec          *ir)
 376 {
 377     assert(ir);
 378
 379     /* override with anything else than the default -2 */
 380     if (nsteps_cmdline > -2)
 381     {
 382         char sbuf_steps[STEPSTRSIZE];
 383         char sbuf_msg[STRLEN];
 384
 385         ir->nsteps = nsteps_cmdline;
 386         if (EI_DYNAMICS(ir->eI) && nsteps_cmdline != -1)
 387         {
 388             sprintf(sbuf_msg, "Overriding nsteps with value passed on the command line: %s steps, %.3g ps",
 389                     gmx_step_str(nsteps_cmdline, sbuf_steps),
 390                     fabs(nsteps_cmdline*ir->delta_t));
 391         }
 392         else
 393         {
 394             sprintf(sbuf_msg, "Overriding nsteps with value passed on the command line: %s steps",
 395                     gmx_step_str(nsteps_cmdline, sbuf_steps));
 396         }
 397
 398         GMX_LOG(mdlog.warning).asParagraph().appendText(sbuf_msg);
 399     }
 400     else if (nsteps_cmdline < -2)
 401     {
 402         gmx_fatal(FARGS, "Invalid nsteps value passed on the command line: %" PRId64,
 403                   nsteps_cmdline);
 404     }
 405     /* Do nothing if nsteps_cmdline == -2 */
 406 }
 407
 408 namespace gmx
 409 {
 410
 411 /*! \brief Return whether GPU acceleration of nonbondeds is supported with the given settings.
 412  *
 413  * If not, and if a warning may be issued, logs a warning about
 414  * falling back to CPU code. With thread-MPI, only the first
 415  * call to this function should have \c issueWarning true. */
 416 static bool gpuAccelerationOfNonbondedIsUseful(const MDLogger   &mdlog,
 417                                                const t_inputrec &ir,
 418                                                bool              issueWarning)
 419 {
 420     bool        gpuIsUseful = true;
 421     std::string warning;
 422
 423     if (ir.opts.ngener - ir.nwall > 1)
 424     {
 425         /* The GPU code does not support more than one energy group.
 426          * If the user requested GPUs explicitly, a fatal error is given later.
 427          */
 428         gpuIsUseful = false;
 429         warning     =
 430             "Multiple energy groups is not implemented for GPUs, falling back to the CPU. "
 431             "For better performance, run on the GPU without energy groups and then do "
 432             "gmx mdrun -rerun option on the trajectory with an energy group .tpr file.";
 433     }
 434
 435     if (EI_TPI(ir.eI))
 436     {
 437         gpuIsUseful = false;
 438         warning     = "TPI is not implemented for GPUs.";
 439     }
 440
 441     if (!gpuIsUseful && issueWarning)
 442     {
 443         GMX_LOG(mdlog.warning).asParagraph().appendText(warning);
 444     }
 445
 446     return gpuIsUseful;
 447 }
 448
 449 //! Initializes the logger for mdrun.
 450 static gmx::LoggerOwner buildLogger(FILE *fplog, const t_commrec *cr)
 451 {
 452     gmx::LoggerBuilder builder;
 453     if (fplog != nullptr)
 454     {
 455         builder.addTargetFile(gmx::MDLogger::LogLevel::Info, fplog);
 456     }
 457     if (cr == nullptr || SIMMASTER(cr))
 458     {
 459         builder.addTargetStream(gmx::MDLogger::LogLevel::Warning,
 460                                 &gmx::TextOutputFile::standardError());
 461     }
 462     return builder.build();
 463 }
 464
 465 //! Make a TaskTarget from an mdrun argument string.
 466 static TaskTarget findTaskTarget(const char *optionString)
 467 {
 468     TaskTarget returnValue = TaskTarget::Auto;
 469
 470     if (strncmp(optionString, "auto", 3) == 0)
 471     {
 472         returnValue = TaskTarget::Auto;
 473     }
 474     else if (strncmp(optionString, "cpu", 3) == 0)
 475     {
 476         returnValue = TaskTarget::Cpu;
 477     }
 478     else if (strncmp(optionString, "gpu", 3) == 0)
 479     {
 480         returnValue = TaskTarget::Gpu;
 481     }
 482     else
 483     {
 484         GMX_ASSERT(false, "Option string should have been checked for sanity already");
 485     }
 486
 487     return returnValue;
 488 }
 489
 490 //! Finish run, aggregate data to print performance info.
 491 static void finish_run(FILE *fplog,
 492                        const gmx::MDLogger &mdlog,
 493                        const t_commrec *cr,
 494                        const t_inputrec *inputrec,
 495                        t_nrnb nrnb[], gmx_wallcycle_t wcycle,
 496                        gmx_walltime_accounting_t walltime_accounting,
 497                        nonbonded_verlet_t *nbv,
 498                        const gmx_pme_t *pme,
 499                        gmx_bool bWriteStat)
 500 {
 501     double  delta_t  = 0;
 502     double  nbfs     = 0, mflop = 0;
 503     double  elapsed_time,
 504             elapsed_time_over_all_ranks,
 505             elapsed_time_over_all_threads,
 506             elapsed_time_over_all_threads_over_all_ranks;
 507     /* Control whether it is valid to print a report. Only the
 508        simulation master may print, but it should not do so if the run
 509        terminated e.g. before a scheduled reset step. This is
 510        complicated by the fact that PME ranks are unaware of the
 511        reason why they were sent a pmerecvqxFINISH. To avoid
 512        communication deadlocks, we always do the communication for the
 513        report, even if we've decided not to write the report, because
 514        how long it takes to finish the run is not important when we've
 515        decided not to report on the simulation performance.
 516
 517        Further, we only report performance for dynamical integrators,
 518        because those are the only ones for which we plan to
 519        consider doing any optimizations. */
 520     bool printReport = EI_DYNAMICS(inputrec->eI) && SIMMASTER(cr);
 521
 522     if (printReport && !walltime_accounting_get_valid_finish(walltime_accounting))
 523     {
 524         GMX_LOG(mdlog.warning).asParagraph().appendText("Simulation ended prematurely, no performance report will be written.");
 525         printReport = false;
 526     }
 527
 528     t_nrnb                  *nrnb_tot;
 529     std::unique_ptr<t_nrnb>  nrnbTotalStorage;
 530     if (cr->nnodes > 1)
 531     {
 532         nrnbTotalStorage = std::make_unique<t_nrnb>();
 533         nrnb_tot         = nrnbTotalStorage.get();
 534 #if GMX_MPI
 535         MPI_Allreduce(nrnb->n, nrnb_tot->n, eNRNB, MPI_DOUBLE, MPI_SUM,
 536                       cr->mpi_comm_mysim);
 537 #endif
 538     }
 539     else
 540     {
 541         nrnb_tot = nrnb;
 542     }
 543
 544     elapsed_time                  = walltime_accounting_get_time_since_reset(walltime_accounting);
 545     elapsed_time_over_all_threads = walltime_accounting_get_time_since_reset_over_all_threads(walltime_accounting);
 546     if (cr->nnodes > 1)
 547     {
 548 #if GMX_MPI
 549         /* reduce elapsed_time over all MPI ranks in the current simulation */
 550         MPI_Allreduce(&elapsed_time,
 551                       &elapsed_time_over_all_ranks,
 552                       1, MPI_DOUBLE, MPI_SUM,
 553                       cr->mpi_comm_mysim);
 554         elapsed_time_over_all_ranks /= cr->nnodes;
 555         /* Reduce elapsed_time_over_all_threads over all MPI ranks in the
 556          * current simulation. */
 557         MPI_Allreduce(&elapsed_time_over_all_threads,
 558                       &elapsed_time_over_all_threads_over_all_ranks,
 559                       1, MPI_DOUBLE, MPI_SUM,
 560                       cr->mpi_comm_mysim);
 561 #endif
 562     }
 563     else
 564     {
 565         elapsed_time_over_all_ranks                  = elapsed_time;
 566         elapsed_time_over_all_threads_over_all_ranks = elapsed_time_over_all_threads;
 567     }
 568
 569     if (printReport)
 570     {
 571         print_flop(fplog, nrnb_tot, &nbfs, &mflop);
 572     }
 573
 574     if (thisRankHasDuty(cr, DUTY_PP) && DOMAINDECOMP(cr))
 575     {
 576         print_dd_statistics(cr, inputrec, fplog);
 577     }
 578
 579     /* TODO Move the responsibility for any scaling by thread counts
 580      * to the code that handled the thread region, so that there's a
 581      * mechanism to keep cycle counting working during the transition
 582      * to task parallelism. */
 583     int nthreads_pp  = gmx_omp_nthreads_get(emntNonbonded);
 584     int nthreads_pme = gmx_omp_nthreads_get(emntPME);
 585     wallcycle_scale_by_num_threads(wcycle, thisRankHasDuty(cr, DUTY_PME) && !thisRankHasDuty(cr, DUTY_PP), nthreads_pp, nthreads_pme);
 586     auto cycle_sum(wallcycle_sum(cr, wcycle));
 587
 588     if (printReport)
 589     {
 590         auto                    nbnxn_gpu_timings = (nbv != nullptr && nbv->useGpu()) ? Nbnxm::gpu_get_timings(nbv->gpu_nbv) : nullptr;
 591         gmx_wallclock_gpu_pme_t pme_gpu_timings   = {};
 592
 593         if (pme_gpu_task_enabled(pme))
 594         {
 595             pme_gpu_get_timings(pme, &pme_gpu_timings);
 596         }
 597         wallcycle_print(fplog, mdlog, cr->nnodes, cr->npmenodes, nthreads_pp, nthreads_pme,
 598                         elapsed_time_over_all_ranks,
 599                         wcycle, cycle_sum,
 600                         nbnxn_gpu_timings,
 601                         &pme_gpu_timings);
 602
 603         if (EI_DYNAMICS(inputrec->eI))
 604         {
 605             delta_t = inputrec->delta_t;
 606         }
 607
 608         if (fplog)
 609         {
 610             print_perf(fplog, elapsed_time_over_all_threads_over_all_ranks,
 611                        elapsed_time_over_all_ranks,
 612                        walltime_accounting_get_nsteps_done_since_reset(walltime_accounting),
 613                        delta_t, nbfs, mflop);
 614         }
 615         if (bWriteStat)
 616         {
 617             print_perf(stderr, elapsed_time_over_all_threads_over_all_ranks,
 618                        elapsed_time_over_all_ranks,
 619                        walltime_accounting_get_nsteps_done_since_reset(walltime_accounting),
 620                        delta_t, nbfs, mflop);
 621         }
 622     }
 623 }
 624
 625 int Mdrunner::mdrunner()
 626 {
 627     matrix                    box;
 628     t_forcerec               *fr               = nullptr;
 629     t_fcdata                 *fcd              = nullptr;
 630     real                      ewaldcoeff_q     = 0;
 631     real                      ewaldcoeff_lj    = 0;
 632     int                       nChargePerturbed = -1, nTypePerturbed = 0;
 633     gmx_wallcycle_t           wcycle;
 634     gmx_walltime_accounting_t walltime_accounting = nullptr;
 635     gmx_membed_t *            membed              = nullptr;
 636     gmx_hw_info_t            *hwinfo              = nullptr;
 637
 638     /* CAUTION: threads may be started later on in this function, so
 639        cr doesn't reflect the final parallel state right now */
 640     gmx_mtop_t                      mtop;
 641
 642     bool doMembed = opt2bSet("-membed", filenames.size(), filenames.data());
 643     bool doRerun  = mdrunOptions.rerun;
 644
 645     // Handle task-assignment related user options.
 646     EmulateGpuNonbonded emulateGpuNonbonded = (getenv("GMX_EMULATE_GPU") != nullptr ?
 647                                                EmulateGpuNonbonded::Yes : EmulateGpuNonbonded::No);
 648
 649     std::vector<int> userGpuTaskAssignment;
 650     try
 651     {
 652         userGpuTaskAssignment = parseUserTaskAssignmentString(hw_opt.userGpuTaskAssignment);
 653     }
 654     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
 655     auto       nonbondedTarget = findTaskTarget(nbpu_opt);
 656     auto       pmeTarget       = findTaskTarget(pme_opt);
 657     auto       pmeFftTarget    = findTaskTarget(pme_fft_opt);
 658     auto       bondedTarget    = findTaskTarget(bonded_opt);
 659     auto       updateTarget    = findTaskTarget(update_opt);
 660     PmeRunMode pmeRunMode      = PmeRunMode::None;
 661
 662     // Here we assume that SIMMASTER(cr) does not change even after the
 663     // threads are started.
 664
 665     FILE *fplog = nullptr;
 666     // If we are appending, we don't write log output because we need
 667     // to check that the old log file matches what the checkpoint file
 668     // expects. Otherwise, we should start to write log output now if
 669     // there is a file ready for it.
 670     if (logFileHandle != nullptr && startingBehavior != StartingBehavior::RestartWithAppending)
 671     {
 672         fplog = gmx_fio_getfp(logFileHandle);
 673     }
 674     gmx::LoggerOwner logOwner(buildLogger(fplog, cr));
 675     gmx::MDLogger    mdlog(logOwner.logger());
 676
 677     // report any development features that may be enabled by environment variables
 678     manageDevelopmentFeatures(mdlog);
 679
 680     // With thread-MPI, the communicator changes after threads are
 681     // launched, so this is rebuilt for the master rank at that
 682     // time. The non-master ranks are fine to keep the one made here.
 683     PhysicalNodeCommunicator physicalNodeComm(MPI_COMM_WORLD, gmx_physicalnode_id_hash());
 684     hwinfo = gmx_detect_hardware(mdlog, physicalNodeComm);
 685
 686     gmx_print_detected_hardware(fplog, isMasterSimMasterRank(ms, MASTER(cr)), mdlog, hwinfo);
 687
 688     std::vector<int> gpuIdsToUse = makeGpuIdsToUse(hwinfo->gpu_info, hw_opt.gpuIdsAvailable);
 689
 690     // Print citation requests after all software/hardware printing
 691     pleaseCiteGromacs(fplog);
 692
 693     // TODO Replace this by unique_ptr once t_inputrec is C++
 694     t_inputrec               inputrecInstance;
 695     t_inputrec              *inputrec = nullptr;
 696     std::unique_ptr<t_state> globalState;
 697
 698     if (SIMMASTER(cr))
 699     {
 700         /* Only the master rank has the global state */
 701         globalState = std::make_unique<t_state>();
 702
 703         /* Read (nearly) all data required for the simulation */
 704         read_tpx_state(ftp2fn(efTPR, filenames.size(), filenames.data()),
 705                        &inputrecInstance, globalState.get(), &mtop);
 706         inputrec = &inputrecInstance;
 707     }
 708
 709     /* Check and update the hardware options for internal consistency */
 710     checkAndUpdateHardwareOptions(mdlog, &hw_opt, SIMMASTER(cr), domdecOptions.numPmeRanks, inputrec);
 711
 712     if (GMX_THREAD_MPI && SIMMASTER(cr))
 713     {
 714         bool useGpuForNonbonded = false;
 715         bool useGpuForPme       = false;
 716         try
 717         {
 718             GMX_RELEASE_ASSERT(inputrec != nullptr, "Keep the compiler happy");
 719
 720             // If the user specified the number of ranks, then we must
 721             // respect that, but in default mode, we need to allow for
 722             // the number of GPUs to choose the number of ranks.
 723             auto canUseGpuForNonbonded = buildSupportsNonbondedOnGpu(nullptr);
 724             useGpuForNonbonded = decideWhetherToUseGpusForNonbondedWithThreadMpi
 725                     (nonbondedTarget, gpuIdsToUse, userGpuTaskAssignment, emulateGpuNonbonded,
 726                     canUseGpuForNonbonded,
 727                     gpuAccelerationOfNonbondedIsUseful(mdlog, *inputrec, GMX_THREAD_MPI),
 728                     hw_opt.nthreads_tmpi);
 729             useGpuForPme = decideWhetherToUseGpusForPmeWithThreadMpi
 730                     (useGpuForNonbonded, pmeTarget, gpuIdsToUse, userGpuTaskAssignment,
 731                     *hwinfo, *inputrec, mtop, hw_opt.nthreads_tmpi, domdecOptions.numPmeRanks);
 732
 733         }
 734         GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
 735
 736         /* Determine how many thread-MPI ranks to start.
 737          *
 738          * TODO Over-writing the user-supplied value here does
 739          * prevent any possible subsequent checks from working
 740          * correctly. */
 741         hw_opt.nthreads_tmpi = get_nthreads_mpi(hwinfo,
 742                                                 &hw_opt,
 743                                                 gpuIdsToUse,
 744                                                 useGpuForNonbonded,
 745                                                 useGpuForPme,
 746                                                 inputrec, &mtop,
 747                                                 mdlog,
 748                                                 doMembed);
 749
 750         // Now start the threads for thread MPI.
 751         cr = spawnThreads(hw_opt.nthreads_tmpi);
 752         /* The main thread continues here with a new cr. We don't deallocate
 753            the old cr because other threads may still be reading it. */
 754         // TODO Both master and spawned threads call dup_tfn and
 755         // reinitialize_commrec_for_this_thread. Find a way to express
 756         // this better.
 757         physicalNodeComm = PhysicalNodeCommunicator(MPI_COMM_WORLD, gmx_physicalnode_id_hash());
 758     }
 759     // END OF CAUTION: cr and physicalNodeComm are now reliable
 760
 761     if (PAR(cr))
 762     {
 763         /* now broadcast everything to the non-master nodes/threads: */
 764         if (!SIMMASTER(cr))
 765         {
 766             inputrec = &inputrecInstance;
 767         }
 768         init_parallel(cr, inputrec, &mtop);
 769     }
 770     GMX_RELEASE_ASSERT(inputrec != nullptr, "All range should have a valid inputrec now");
 771
 772     // Now each rank knows the inputrec that SIMMASTER read and used,
 773     // and (if applicable) cr->nnodes has been assigned the number of
 774     // thread-MPI ranks that have been chosen. The ranks can now all
 775     // run the task-deciding functions and will agree on the result
 776     // without needing to communicate.
 777     //
 778     // TODO Should we do the communication in debug mode to support
 779     // having an assertion?
 780     //
 781     // Note that these variables describe only their own node.
 782     //
 783     // Note that when bonded interactions run on a GPU they always run
 784     // alongside a nonbonded task, so do not influence task assignment
 785     // even though they affect the force calculation workload.
 786     bool useGpuForNonbonded     = false;
 787     bool useGpuForPme           = false;
 788     bool useGpuForBonded        = false;
 789     bool useGpuForUpdate        = false;
 790     bool gpusWereDetected       = hwinfo->ngpu_compatible_tot > 0;
 791     try
 792     {
 793         // It's possible that there are different numbers of GPUs on
 794         // different nodes, which is the user's responsibilty to
 795         // handle. If unsuitable, we will notice that during task
 796         // assignment.
 797         auto canUseGpuForNonbonded = buildSupportsNonbondedOnGpu(nullptr);
 798         useGpuForNonbonded = decideWhetherToUseGpusForNonbonded(nonbondedTarget, userGpuTaskAssignment,
 799                                                                 emulateGpuNonbonded,
 800                                                                 canUseGpuForNonbonded,
 801                                                                 gpuAccelerationOfNonbondedIsUseful(mdlog, *inputrec, !GMX_THREAD_MPI),
 802                                                                 gpusWereDetected);
 803         useGpuForPme = decideWhetherToUseGpusForPme(useGpuForNonbonded, pmeTarget, userGpuTaskAssignment,
 804                                                     *hwinfo, *inputrec, mtop,
 805                                                     cr->nnodes, domdecOptions.numPmeRanks,
 806                                                     gpusWereDetected);
 807         auto canUseGpuForBonded = buildSupportsGpuBondeds(nullptr) && inputSupportsGpuBondeds(*inputrec, mtop, nullptr);
 808         useGpuForBonded =
 809             decideWhetherToUseGpusForBonded(useGpuForNonbonded, useGpuForPme,
 810                                             bondedTarget, canUseGpuForBonded,
 811                                             EVDW_PME(inputrec->vdwtype),
 812                                             EEL_PME_EWALD(inputrec->coulombtype),
 813                                             domdecOptions.numPmeRanks, gpusWereDetected);
 814
 815         pmeRunMode   = (useGpuForPme ? PmeRunMode::GPU : PmeRunMode::CPU);
 816         if (pmeRunMode == PmeRunMode::GPU)
 817         {
 818             if (pmeFftTarget == TaskTarget::Cpu)
 819             {
 820                 pmeRunMode = PmeRunMode::Mixed;
 821             }
 822         }
 823         else if (pmeFftTarget == TaskTarget::Gpu)
 824         {
 825             gmx_fatal(FARGS, "Assigning FFTs to GPU requires PME to be assigned to GPU as well. With PME on CPU you should not be using -pmefft.");
 826         }
 827     }
 828     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
 829
 830     // Build restraints.
 831     // TODO: hide restraint implementation details from Mdrunner.
 832     // There is nothing unique about restraints at this point as far as the
 833     // Mdrunner is concerned. The Mdrunner should just be getting a sequence of
 834     // factory functions from the SimulationContext on which to call mdModules_->add().
 835     // TODO: capture all restraints into a single RestraintModule, passed to the runner builder.
 836     for (auto && restraint : restraintManager_->getRestraints())
 837     {
 838         auto module = RestraintMDModule::create(restraint,
 839                                                 restraint->sites());
 840         mdModules_->add(std::move(module));
 841     }
 842
 843     // TODO: Error handling
 844     mdModules_->assignOptionsToModules(*inputrec->params, nullptr);
 845     const auto &mdModulesNotifier = mdModules_->notifier().notifier_;
 846
 847     if (inputrec->internalParameters != nullptr)
 848     {
 849         mdModulesNotifier.notify(*inputrec->internalParameters);
 850     }
 851
 852     if (fplog != nullptr)
 853     {
 854         pr_inputrec(fplog, 0, "Input Parameters", inputrec, FALSE);
 855         fprintf(fplog, "\n");
 856     }
 857
 858     if (SIMMASTER(cr))
 859     {
 860         /* In rerun, set velocities to zero if present */
 861         if (doRerun && ((globalState->flags & (1 << estV)) != 0))
 862         {
 863             // rerun does not use velocities
 864             GMX_LOG(mdlog.info).asParagraph().appendText(
 865                     "Rerun trajectory contains velocities. Rerun does only evaluate "
 866                     "potential energy and forces. The velocities will be ignored.");
 867             for (int i = 0; i < globalState->natoms; i++)
 868             {
 869                 clear_rvec(globalState->v[i]);
 870             }
 871             globalState->flags &= ~(1 << estV);
 872         }
 873
 874         /* now make sure the state is initialized and propagated */
 875         set_state_entries(globalState.get(), inputrec);
 876     }
 877
 878     /* NM and TPI parallelize over force/energy calculations, not atoms,
 879      * so we need to initialize and broadcast the global state.
 880      */
 881     if (inputrec->eI == eiNM || inputrec->eI == eiTPI)
 882     {
 883         if (!MASTER(cr))
 884         {
 885             globalState = std::make_unique<t_state>();
 886         }
 887         broadcastStateWithoutDynamics(cr, globalState.get());
 888     }
 889
 890     /* A parallel command line option consistency check that we can
 891        only do after any threads have started. */
 892     if (!PAR(cr) && (domdecOptions.numCells[XX] > 1 ||
 893                      domdecOptions.numCells[YY] > 1 ||
 894                      domdecOptions.numCells[ZZ] > 1 ||
 895                      domdecOptions.numPmeRanks > 0))
 896     {
 897         gmx_fatal(FARGS,
 898                   "The -dd or -npme option request a parallel simulation, "
 899 #if !GMX_MPI
 900                   "but %s was compiled without threads or MPI enabled", output_env_get_program_display_name(oenv));
 901 #else
 902 #if GMX_THREAD_MPI
 903                   "but the number of MPI-threads (option -ntmpi) is not set or is 1");
 904 #else
 905                   "but %s was not started through mpirun/mpiexec or only one rank was requested through mpirun/mpiexec", output_env_get_program_display_name(oenv));
 906 #endif
 907 #endif
 908     }
 909
 910     if (doRerun &&
 911         (EI_ENERGY_MINIMIZATION(inputrec->eI) || eiNM == inputrec->eI))
 912     {
 913         gmx_fatal(FARGS, "The .mdp file specified an energy mininization or normal mode algorithm, and these are not compatible with mdrun -rerun");
 914     }
 915
 916     if (!(EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype)))
 917     {
 918         if (domdecOptions.numPmeRanks > 0)
 919         {
 920             gmx_fatal_collective(FARGS, cr->mpi_comm_mysim, MASTER(cr),
 921                                  "PME-only ranks are requested, but the system does not use PME for electrostatics or LJ");
 922         }
 923
 924         domdecOptions.numPmeRanks = 0;
 925     }
 926
 927     if (useGpuForNonbonded && domdecOptions.numPmeRanks < 0)
 928     {
 929         /* With NB GPUs we don't automatically use PME-only CPU ranks. PME ranks can
 930          * improve performance with many threads per GPU, since our OpenMP
 931          * scaling is bad, but it's difficult to automate the setup.
 932          */
 933         domdecOptions.numPmeRanks = 0;
 934     }
 935     if (useGpuForPme)
 936     {
 937         if (domdecOptions.numPmeRanks < 0)
 938         {
 939             domdecOptions.numPmeRanks = 0;
 940             // TODO possibly print a note that one can opt-in for a separate PME GPU rank?
 941         }
 942         else
 943         {
 944             GMX_RELEASE_ASSERT(domdecOptions.numPmeRanks <= 1, "PME GPU decomposition is not supported");
 945         }
 946     }
 947
 948 #if GMX_FAHCORE
 949     if (MASTER(cr))
 950     {
 951         fcRegisterSteps(inputrec->nsteps, inputrec->init_step);
 952     }
 953 #endif
 954
 955     /* NMR restraints must be initialized before load_checkpoint,
 956      * since with time averaging the history is added to t_state.
 957      * For proper consistency check we therefore need to extend
 958      * t_state here.
 959      * So the PME-only nodes (if present) will also initialize
 960      * the distance restraints.
 961      */
 962     snew(fcd, 1);
 963
 964     /* This needs to be called before read_checkpoint to extend the state */
 965     init_disres(fplog, &mtop, inputrec, cr, ms, fcd, globalState.get(), replExParams.exchangeInterval > 0);
 966
 967     init_orires(fplog, &mtop, inputrec, cr, ms, globalState.get(), &(fcd->orires));
 968
 969     auto                 deform = prepareBoxDeformation(globalState->box, cr, *inputrec);
 970
 971     ObservablesHistory   observablesHistory = {};
 972
 973     if (startingBehavior != StartingBehavior::NewSimulation)
 974     {
 975         /* Check if checkpoint file exists before doing continuation.
 976          * This way we can use identical input options for the first and subsequent runs...
 977          */
 978         if (mdrunOptions.numStepsCommandline > -2)
 979         {
 980             /* Temporarily set the number of steps to unmlimited to avoid
 981              * triggering the nsteps check in load_checkpoint().
 982              * This hack will go away soon when the -nsteps option is removed.
 983              */
 984             inputrec->nsteps = -1;
 985         }
 986
 987         load_checkpoint(opt2fn_master("-cpi", filenames.size(), filenames.data(), cr),
 988                         logFileHandle,
 989                         cr, domdecOptions.numCells,
 990                         inputrec, globalState.get(),
 991                         &observablesHistory,
 992                         mdrunOptions.reproducible, mdModules_->notifier());
 993
 994         if (startingBehavior == StartingBehavior::RestartWithAppending && logFileHandle)
 995         {
 996             // Now we can start normal logging to the truncated log file.
 997             fplog    = gmx_fio_getfp(logFileHandle);
 998             prepareLogAppending(fplog);
 999             logOwner = buildLogger(fplog, cr);
1000             mdlog    = logOwner.logger();
1001         }
1002     }
1003
1004     if (mdrunOptions.numStepsCommandline > -2)
1005     {
1006         GMX_LOG(mdlog.info).asParagraph().
1007             appendText("The -nsteps functionality is deprecated, and may be removed in a future version. "
1008                        "Consider using gmx convert-tpr -nsteps or changing the appropriate .mdp file field.");
1009     }
1010     /* override nsteps with value set on the commamdline */
1011     override_nsteps_cmdline(mdlog, mdrunOptions.numStepsCommandline, inputrec);
1012
1013     if (SIMMASTER(cr))
1014     {
1015         copy_mat(globalState->box, box);
1016     }
1017
1018     if (PAR(cr))
1019     {
1020         gmx_bcast(sizeof(box), box, cr);
1021     }
1022
1023     if (inputrec->cutoff_scheme != ecutsVERLET)
1024     {
1025         gmx_fatal(FARGS, "This group-scheme .tpr file can no longer be run by mdrun. Please update to the Verlet scheme, or use an earlier version of GROMACS if necessary.");
1026     }
1027     /* Update rlist and nstlist. */
1028     prepare_verlet_scheme(fplog, cr, inputrec, nstlist_cmdline, &mtop, box,
1029                           useGpuForNonbonded || (emulateGpuNonbonded == EmulateGpuNonbonded::Yes), *hwinfo->cpuInfo);
1030
1031     LocalAtomSetManager atomSets;
1032     if (PAR(cr) && !(EI_TPI(inputrec->eI) ||
1033                      inputrec->eI == eiNM))
1034     {
1035         cr->dd = init_domain_decomposition(mdlog, cr, domdecOptions, mdrunOptions,
1036                                            &mtop, inputrec,
1037                                            box, positionsFromStatePointer(globalState.get()),
1038                                            &atomSets);
1039         // Note that local state still does not exist yet.
1040     }
1041     else
1042     {
1043         /* PME, if used, is done on all nodes with 1D decomposition */
1044         cr->npmenodes = 0;
1045         cr->duty      = (DUTY_PP | DUTY_PME);
1046
1047         if (inputrec->ePBC == epbcSCREW)
1048         {
1049             gmx_fatal(FARGS,
1050                       "pbc=screw is only implemented with domain decomposition");
1051         }
1052     }
1053
1054     if (PAR(cr))
1055     {
1056         /* After possible communicator splitting in make_dd_communicators.
1057          * we can set up the intra/inter node communication.
1058          */
1059         gmx_setup_nodecomm(fplog, cr);
1060     }
1061
1062 #if GMX_MPI
1063     if (isMultiSim(ms))
1064     {
1065         GMX_LOG(mdlog.warning).asParagraph().appendTextFormatted(
1066                 "This is simulation %d out of %d running as a composite GROMACS\n"
1067                 "multi-simulation job. Setup for this simulation:\n",
1068                 ms->sim, ms->nsim);
1069     }
1070     GMX_LOG(mdlog.warning).appendTextFormatted(
1071             "Using %d MPI %s\n",
1072             cr->nnodes,
1073 #if GMX_THREAD_MPI
1074             cr->nnodes == 1 ? "thread" : "threads"
1075 #else
1076             cr->nnodes == 1 ? "process" : "processes"
1077 #endif
1078             );
1079     fflush(stderr);
1080 #endif
1081
1082     // If mdrun -pin auto honors any affinity setting that already
1083     // exists. If so, it is nice to provide feedback about whether
1084     // that existing affinity setting was from OpenMP or something
1085     // else, so we run this code both before and after we initialize
1086     // the OpenMP support.
1087     gmx_check_thread_affinity_set(mdlog,
1088                                   &hw_opt, hwinfo->nthreads_hw_avail, FALSE);
1089     /* Check and update the number of OpenMP threads requested */
1090     checkAndUpdateRequestedNumOpenmpThreads(&hw_opt, *hwinfo, cr, ms, physicalNodeComm.size_,
1091                                             pmeRunMode, mtop, *inputrec);
1092
1093     gmx_omp_nthreads_init(mdlog, cr,
1094                           hwinfo->nthreads_hw_avail,
1095                           physicalNodeComm.size_,
1096                           hw_opt.nthreads_omp,
1097                           hw_opt.nthreads_omp_pme,
1098                           !thisRankHasDuty(cr, DUTY_PP));
1099
1100     // Enable FP exception detection, but not in
1101     // Release mode and not for compilers with known buggy FP
1102     // exception support (clang with any optimization) or suspected
1103     // buggy FP exception support (gcc 7.* with optimization).
1104 #if !defined NDEBUG && \
1105     !((defined __clang__ || (defined(__GNUC__) && !defined(__ICC) && __GNUC__ == 7)) \
1106     && defined __OPTIMIZE__)
1107     const bool bEnableFPE = true;
1108 #else
1109     const bool bEnableFPE = false;
1110 #endif
1111     //FIXME - reconcile with gmx_feenableexcept() call from CommandLineModuleManager::run()
1112     if (bEnableFPE)
1113     {
1114         gmx_feenableexcept();
1115     }
1116
1117     // Build a data structure that expresses which kinds of non-bonded
1118     // task are handled by this rank.
1119     //
1120     // TODO Later, this might become a loop over all registered modules
1121     // relevant to the mdp inputs, to find those that have such tasks.
1122     //
1123     // TODO This could move before init_domain_decomposition() as part
1124     // of refactoring that separates the responsibility for duty
1125     // assignment from setup for communication between tasks, and
1126     // setup for tasks handled with a domain (ie including short-ranged
1127     // tasks, bonded tasks, etc.).
1128     //
1129     // Note that in general useGpuForNonbonded, etc. can have a value
1130     // that is inconsistent with the presence of actual GPUs on any
1131     // rank, and that is not known to be a problem until the
1132     // duty of the ranks on a node become known.
1133     //
1134     // TODO Later we might need the concept of computeTasksOnThisRank,
1135     // from which we construct gpuTasksOnThisRank.
1136     //
1137     // Currently the DD code assigns duty to ranks that can
1138     // include PP work that currently can be executed on a single
1139     // GPU, if present and compatible.  This has to be coordinated
1140     // across PP ranks on a node, with possible multiple devices
1141     // or sharing devices on a node, either from the user
1142     // selection, or automatically.
1143     auto                 haveGpus = !gpuIdsToUse.empty();
1144     std::vector<GpuTask> gpuTasksOnThisRank;
1145     if (thisRankHasDuty(cr, DUTY_PP))
1146     {
1147         if (useGpuForNonbonded)
1148         {
1149             // Note that any bonded tasks on a GPU always accompany a
1150             // non-bonded task.
1151             if (haveGpus)
1152             {
1153                 gpuTasksOnThisRank.push_back(GpuTask::Nonbonded);
1154             }
1155             else if (nonbondedTarget == TaskTarget::Gpu)
1156             {
1157                 gmx_fatal(FARGS, "Cannot run short-ranged nonbonded interactions on a GPU because no GPU is detected.");
1158             }
1159             else if (bondedTarget == TaskTarget::Gpu)
1160             {
1161                 gmx_fatal(FARGS, "Cannot run bonded interactions on a GPU because no GPU is detected.");
1162             }
1163         }
1164     }
1165     // TODO cr->duty & DUTY_PME should imply that a PME algorithm is active, but currently does not.
1166     if (EEL_PME(inputrec->coulombtype) && (thisRankHasDuty(cr, DUTY_PME)))
1167     {
1168         if (useGpuForPme)
1169         {
1170             if (haveGpus)
1171             {
1172                 gpuTasksOnThisRank.push_back(GpuTask::Pme);
1173             }
1174             else if (pmeTarget == TaskTarget::Gpu)
1175             {
1176                 gmx_fatal(FARGS, "Cannot run PME on a GPU because no GPU is detected.");
1177             }
1178         }
1179     }
1180
1181     GpuTaskAssignment gpuTaskAssignment;
1182     try
1183     {
1184         // Produce the task assignment for this rank.
1185         gpuTaskAssignment = runTaskAssignment(gpuIdsToUse, userGpuTaskAssignment, *hwinfo,
1186                                               mdlog, cr, ms, physicalNodeComm, gpuTasksOnThisRank,
1187                                               useGpuForBonded, pmeRunMode);
1188     }
1189     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1190
1191     /* Prevent other ranks from continuing after an issue was found
1192      * and reported as a fatal error.
1193      *
1194      * TODO This function implements a barrier so that MPI runtimes
1195      * can organize an orderly shutdown if one of the ranks has had to
1196      * issue a fatal error in various code already run. When we have
1197      * MPI-aware error handling and reporting, this should be
1198      * improved. */
1199 #if GMX_MPI
1200     if (PAR(cr))
1201     {
1202         MPI_Barrier(cr->mpi_comm_mysim);
1203     }
1204     if (isMultiSim(ms))
1205     {
1206         if (SIMMASTER(cr))
1207         {
1208             MPI_Barrier(ms->mpi_comm_masters);
1209         }
1210         /* We need another barrier to prevent non-master ranks from contiuing
1211          * when an error occured in a different simulation.
1212          */
1213         MPI_Barrier(cr->mpi_comm_mysim);
1214     }
1215 #endif
1216
1217     /* Now that we know the setup is consistent, check for efficiency */
1218     check_resource_division_efficiency(hwinfo, !gpuTaskAssignment.empty(), mdrunOptions.ntompOptionIsSet,
1219                                        cr, mdlog);
1220
1221     gmx_device_info_t *nonbondedDeviceInfo = nullptr;
1222
1223     if (thisRankHasDuty(cr, DUTY_PP))
1224     {
1225         // This works because only one task of each type is currently permitted.
1226         auto nbGpuTaskMapping = std::find_if(gpuTaskAssignment.begin(), gpuTaskAssignment.end(),
1227                                              hasTaskType<GpuTask::Nonbonded>);
1228         if (nbGpuTaskMapping != gpuTaskAssignment.end())
1229         {
1230             int nonbondedDeviceId = nbGpuTaskMapping->deviceId_;
1231             nonbondedDeviceInfo = getDeviceInfo(hwinfo->gpu_info, nonbondedDeviceId);
1232             init_gpu(nonbondedDeviceInfo);
1233
1234             if (DOMAINDECOMP(cr))
1235             {
1236                 /* When we share GPUs over ranks, we need to know this for the DLB */
1237                 dd_setup_dlb_resource_sharing(cr, nonbondedDeviceId);
1238             }
1239
1240         }
1241     }
1242
1243     gmx_device_info_t                *pmeDeviceInfo = nullptr;
1244     // Later, this program could contain kernels that might be later
1245     // re-used as auto-tuning progresses, or subsequent simulations
1246     // are invoked.
1247     PmeGpuProgramStorage pmeGpuProgram;
1248     // This works because only one task of each type is currently permitted.
1249     auto                 pmeGpuTaskMapping     = std::find_if(gpuTaskAssignment.begin(), gpuTaskAssignment.end(), hasTaskType<GpuTask::Pme>);
1250     const bool           thisRankHasPmeGpuTask = (pmeGpuTaskMapping != gpuTaskAssignment.end());
1251     if (thisRankHasPmeGpuTask)
1252     {
1253         pmeDeviceInfo = getDeviceInfo(hwinfo->gpu_info, pmeGpuTaskMapping->deviceId_);
1254         init_gpu(pmeDeviceInfo);
1255         pmeGpuProgram = buildPmeGpuProgram(pmeDeviceInfo);
1256     }
1257
1258     /* getting number of PP/PME threads on this MPI / tMPI rank.
1259        PME: env variable should be read only on one node to make sure it is
1260        identical everywhere;
1261      */
1262     const int numThreadsOnThisRank =
1263         thisRankHasDuty(cr, DUTY_PP) ? gmx_omp_nthreads_get(emntNonbonded) : gmx_omp_nthreads_get(emntPME);
1264     checkHardwareOversubscription(numThreadsOnThisRank, cr->nodeid,
1265                                   *hwinfo->hardwareTopology,
1266                                   physicalNodeComm, mdlog);
1267
1268     if (hw_opt.threadAffinity != ThreadAffinity::Off)
1269     {
1270         /* Before setting affinity, check whether the affinity has changed
1271          * - which indicates that probably the OpenMP library has changed it
1272          * since we first checked).
1273          */
1274         gmx_check_thread_affinity_set(mdlog,
1275                                       &hw_opt, hwinfo->nthreads_hw_avail, TRUE);
1276
1277         int numThreadsOnThisNode, intraNodeThreadOffset;
1278         analyzeThreadsOnThisNode(physicalNodeComm, numThreadsOnThisRank, &numThreadsOnThisNode,
1279                                  &intraNodeThreadOffset);
1280
1281         /* Set the CPU affinity */
1282         gmx_set_thread_affinity(mdlog, cr, &hw_opt, *hwinfo->hardwareTopology,
1283                                 numThreadsOnThisRank, numThreadsOnThisNode,
1284                                 intraNodeThreadOffset, nullptr);
1285     }
1286
1287     if (mdrunOptions.timingOptions.resetStep > -1)
1288     {
1289         GMX_LOG(mdlog.info).asParagraph().
1290             appendText("The -resetstep functionality is deprecated, and may be removed in a future version.");
1291     }
1292     wcycle = wallcycle_init(fplog, mdrunOptions.timingOptions.resetStep, cr);
1293
1294     if (PAR(cr))
1295     {
1296         /* Master synchronizes its value of reset_counters with all nodes
1297          * including PME only nodes */
1298         int64_t reset_counters = wcycle_get_reset_counters(wcycle);
1299         gmx_bcast_sim(sizeof(reset_counters), &reset_counters, cr);
1300         wcycle_set_reset_counters(wcycle, reset_counters);
1301     }
1302
1303     // Membrane embedding must be initialized before we call init_forcerec()
1304     if (doMembed)
1305     {
1306         if (MASTER(cr))
1307         {
1308             fprintf(stderr, "Initializing membed");
1309         }
1310         /* Note that membed cannot work in parallel because mtop is
1311          * changed here. Fix this if we ever want to make it run with
1312          * multiple ranks. */
1313         membed = init_membed(fplog, filenames.size(), filenames.data(), &mtop, inputrec, globalState.get(), cr,
1314                              &mdrunOptions
1315                                  .checkpointOptions.period);
1316     }
1317
1318     std::unique_ptr<MDAtoms>     mdAtoms;
1319     std::unique_ptr<gmx_vsite_t> vsite;
1320
1321     t_nrnb nrnb;
1322     if (thisRankHasDuty(cr, DUTY_PP))
1323     {
1324         mdModulesNotifier.notify(*cr);
1325         mdModulesNotifier.notify(&atomSets);
1326         mdModulesNotifier.notify(PeriodicBoundaryConditionType {inputrec->ePBC});
1327         /* Initiate forcerecord */
1328         fr                 = new t_forcerec;
1329         fr->forceProviders = mdModules_->initForceProviders();
1330         init_forcerec(fplog, mdlog, fr, fcd,
1331                       inputrec, &mtop, cr, box,
1332                       opt2fn("-table", filenames.size(), filenames.data()),
1333                       opt2fn("-tablep", filenames.size(), filenames.data()),
1334                       opt2fns("-tableb", filenames.size(), filenames.data()),
1335                       *hwinfo, nonbondedDeviceInfo,
1336                       useGpuForBonded,
1337                       pforce,
1338                       wcycle);
1339
1340         // TODO Move this to happen during domain decomposition setup,
1341         // once stream and event handling works well with that.
1342         // TODO remove need to pass local stream into GPU halo exchange - Redmine #3093
1343         if (havePPDomainDecomposition(cr) && c_enableGpuHaloExchange && useGpuForNonbonded)
1344         {
1345             void *streamLocal                   = Nbnxm::gpu_get_command_stream(fr->nbv->gpu_nbv, Nbnxm::InteractionLocality::NonLocal);
1346             void *streamNonLocal                =
1347                 Nbnxm::gpu_get_command_stream(fr->nbv->gpu_nbv, Nbnxm::InteractionLocality::NonLocal);
1348             void *coordinatesOnDeviceEvent = fr->nbv->get_x_on_device_event();
1349             cr->dd->gpuHaloExchange = std::make_unique<GpuHaloExchange>(cr->dd, cr->mpi_comm_mysim, streamLocal,
1350                                                                         streamNonLocal, coordinatesOnDeviceEvent);
1351         }
1352
1353         /* Initialize the mdAtoms structure.
1354          * mdAtoms is not filled with atom data,
1355          * as this can not be done now with domain decomposition.
1356          */
1357         mdAtoms = makeMDAtoms(fplog, mtop, *inputrec, thisRankHasPmeGpuTask);
1358         if (globalState && thisRankHasPmeGpuTask)
1359         {
1360             // The pinning of coordinates in the global state object works, because we only use
1361             // PME on GPU without DD or on a separate PME rank, and because the local state pointer
1362             // points to the global state object without DD.
1363             // FIXME: MD and EM separately set up the local state - this should happen in the same function,
1364             // which should also perform the pinning.
1365             changePinningPolicy(&globalState->x, pme_get_pinning_policy());
1366         }
1367
1368         /* Initialize the virtual site communication */
1369         vsite = initVsite(mtop, cr);
1370
1371         calc_shifts(box, fr->shift_vec);
1372
1373         /* With periodic molecules the charge groups should be whole at start up
1374          * and the virtual sites should not be far from their proper positions.
1375          */
1376         if (!inputrec->bContinuation && MASTER(cr) &&
1377             !(inputrec->ePBC != epbcNONE && inputrec->bPeriodicMols))
1378         {
1379             /* Make molecules whole at start of run */
1380             if (fr->ePBC != epbcNONE)
1381             {
1382                 do_pbc_first_mtop(fplog, inputrec->ePBC, box, &mtop, globalState->x.rvec_array());
1383             }
1384             if (vsite)
1385             {
1386                 /* Correct initial vsite positions are required
1387                  * for the initial distribution in the domain decomposition
1388                  * and for the initial shell prediction.
1389                  */
1390                 constructVsitesGlobal(mtop, globalState->x);
1391             }
1392         }
1393
1394         if (EEL_PME(fr->ic->eeltype) || EVDW_PME(fr->ic->vdwtype))
1395         {
1396             ewaldcoeff_q  = fr->ic->ewaldcoeff_q;
1397             ewaldcoeff_lj = fr->ic->ewaldcoeff_lj;
1398         }
1399     }
1400     else
1401     {
1402         /* This is a PME only node */
1403
1404         GMX_ASSERT(globalState == nullptr, "We don't need the state on a PME only rank and expect it to be unitialized");
1405
1406         ewaldcoeff_q  = calc_ewaldcoeff_q(inputrec->rcoulomb, inputrec->ewald_rtol);
1407         ewaldcoeff_lj = calc_ewaldcoeff_lj(inputrec->rvdw, inputrec->ewald_rtol_lj);
1408     }
1409
1410     gmx_pme_t *sepPmeData = nullptr;
1411     // This reference hides the fact that PME data is owned by runner on PME-only ranks and by forcerec on other ranks
1412     GMX_ASSERT(thisRankHasDuty(cr, DUTY_PP) == (fr != nullptr), "Double-checking that only PME-only ranks have no forcerec");
1413     gmx_pme_t * &pmedata = fr ? fr->pmedata : sepPmeData;
1414
1415     /* Initiate PME if necessary,
1416      * either on all nodes or on dedicated PME nodes only. */
1417     if (EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype))
1418     {
1419         if (mdAtoms && mdAtoms->mdatoms())
1420         {
1421             nChargePerturbed = mdAtoms->mdatoms()->nChargePerturbed;
1422             if (EVDW_PME(inputrec->vdwtype))
1423             {
1424                 nTypePerturbed   = mdAtoms->mdatoms()->nTypePerturbed;
1425             }
1426         }
1427         if (cr->npmenodes > 0)
1428         {
1429             /* The PME only nodes need to know nChargePerturbed(FEP on Q) and nTypePerturbed(FEP on LJ)*/
1430             gmx_bcast_sim(sizeof(nChargePerturbed), &nChargePerturbed, cr);
1431             gmx_bcast_sim(sizeof(nTypePerturbed), &nTypePerturbed, cr);
1432         }
1433
1434         if (thisRankHasDuty(cr, DUTY_PME))
1435         {
1436             try
1437             {
1438                 pmedata = gmx_pme_init(cr,
1439                                        getNumPmeDomains(cr->dd),
1440                                        inputrec,
1441                                        nChargePerturbed != 0, nTypePerturbed != 0,
1442                                        mdrunOptions.reproducible,
1443                                        ewaldcoeff_q, ewaldcoeff_lj,
1444                                        gmx_omp_nthreads_get(emntPME),
1445                                        pmeRunMode, nullptr,
1446                                        pmeDeviceInfo, pmeGpuProgram.get(), mdlog);
1447             }
1448             GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1449         }
1450     }
1451
1452
1453     if (EI_DYNAMICS(inputrec->eI))
1454     {
1455         /* Turn on signal handling on all nodes */
1456         /*
1457          * (A user signal from the PME nodes (if any)
1458          * is communicated to the PP nodes.
1459          */
1460         signal_handler_install();
1461     }
1462
1463     pull_t *pull_work = nullptr;
1464     if (thisRankHasDuty(cr, DUTY_PP))
1465     {
1466         /* Assumes uniform use of the number of OpenMP threads */
1467         walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(emntDefault));
1468
1469         if (inputrec->bPull)
1470         {
1471             /* Initialize pull code */
1472             pull_work =
1473                 init_pull(fplog, inputrec->pull, inputrec,
1474                           &mtop, cr, &atomSets, inputrec->fepvals->init_lambda);
1475             if (inputrec->pull->bXOutAverage || inputrec->pull->bFOutAverage)
1476             {
1477                 initPullHistory(pull_work, &observablesHistory);
1478             }
1479             if (EI_DYNAMICS(inputrec->eI) && MASTER(cr))
1480             {
1481                 init_pull_output_files(pull_work,
1482                                        filenames.size(), filenames.data(), oenv,
1483                                        startingBehavior);
1484             }
1485         }
1486
1487         std::unique_ptr<EnforcedRotation> enforcedRotation;
1488         if (inputrec->bRot)
1489         {
1490             /* Initialize enforced rotation code */
1491             enforcedRotation = init_rot(fplog,
1492                                         inputrec,
1493                                         filenames.size(),
1494                                         filenames.data(),
1495                                         cr,
1496                                         &atomSets,
1497                                         globalState.get(),
1498                                         &mtop,
1499                                         oenv,
1500                                         mdrunOptions,
1501                                         startingBehavior);
1502         }
1503
1504         t_swap *swap = nullptr;
1505         if (inputrec->eSwapCoords != eswapNO)
1506         {
1507             /* Initialize ion swapping code */
1508             swap = init_swapcoords(fplog, inputrec,
1509                                    opt2fn_master("-swap", filenames.size(), filenames.data(), cr),
1510                                    &mtop, globalState.get(), &observablesHistory,
1511                                    cr, &atomSets, oenv, mdrunOptions,
1512                                    startingBehavior);
1513         }
1514
1515         /* Let makeConstraints know whether we have essential dynamics constraints.
1516          * TODO: inputrec should tell us whether we use an algorithm, not a file option or the checkpoint
1517          */
1518         bool doEssentialDynamics = (opt2fn_null("-ei", filenames.size(), filenames.data()) != nullptr
1519                                     || observablesHistory.edsamHistory);
1520         auto constr              = makeConstraints(mtop, *inputrec, pull_work, doEssentialDynamics,
1521                                                    fplog, *mdAtoms->mdatoms(),
1522                                                    cr, ms, &nrnb, wcycle, fr->bMolPBC);
1523
1524         /* Energy terms and groups */
1525         gmx_enerdata_t enerd(mtop.groups.groups[SimulationAtomGroupType::EnergyOutput].size(), inputrec->fepvals->n_lambda);
1526
1527         /* Kinetic energy data */
1528         gmx_ekindata_t ekind;
1529         init_ekindata(fplog, &mtop, &(inputrec->opts), &ekind, inputrec->cos_accel);
1530
1531         /* Set up interactive MD (IMD) */
1532         auto imdSession = makeImdSession(inputrec, cr, wcycle, &enerd, ms, &mtop, mdlog,
1533                                          MASTER(cr) ? globalState->x.rvec_array() : nullptr,
1534                                          filenames.size(), filenames.data(), oenv, mdrunOptions.imdOptions,
1535                                          startingBehavior);
1536
1537         if (DOMAINDECOMP(cr))
1538         {
1539             GMX_RELEASE_ASSERT(fr, "fr was NULL while cr->duty was DUTY_PP");
1540             /* This call is not included in init_domain_decomposition mainly
1541              * because fr->cginfo_mb is set later.
1542              */
1543             dd_init_bondeds(fplog, cr->dd, &mtop, vsite.get(), inputrec,
1544                             domdecOptions.checkBondedInteractions,
1545                             fr->cginfo_mb);
1546         }
1547
1548         if (updateTarget == TaskTarget::Gpu)
1549         {
1550             if (SIMMASTER(cr))
1551             {
1552                 gmx_fatal(FARGS, "It is currently not possible to redirect the calculation "
1553                           "of update and constraints to the GPU!");
1554             }
1555         }
1556
1557         // Before we start the actual simulator, try if we can run the update task on the GPU.
1558         useGpuForUpdate = decideWhetherToUseGpuForUpdate(DOMAINDECOMP(cr),
1559                                                          useGpuForNonbonded,
1560                                                          updateTarget,
1561                                                          gpusWereDetected,
1562                                                          *inputrec,
1563                                                          *mdAtoms,
1564                                                          doEssentialDynamics,
1565                                                          fcd->orires.nr != 0,
1566                                                          fcd->disres.nsystems != 0);
1567
1568         // TODO This is not the right place to manage the lifetime of
1569         // this data structure, but currently it's the easiest way to
1570         // make it work.
1571         MdrunScheduleWorkload runScheduleWork;
1572
1573         GMX_ASSERT(stopHandlerBuilder_, "Runner must provide StopHandlerBuilder to simulator.");
1574         SimulatorBuilder simulatorBuilder;
1575
1576         // build and run simulator object based on user-input
1577         const bool inputIsCompatibleWithModularSimulator = ModularSimulator::isInputCompatible(
1578                     false,
1579                     inputrec, doRerun, vsite.get(), ms, replExParams,
1580                     fcd, static_cast<int>(filenames.size()), filenames.data(),
1581                     &observablesHistory, membed);
1582         auto simulator = simulatorBuilder.build(
1583                     inputIsCompatibleWithModularSimulator,
1584                     fplog, cr, ms, mdlog, static_cast<int>(filenames.size()), filenames.data(),
1585                     oenv,
1586                     mdrunOptions,
1587                     startingBehavior,
1588                     vsite.get(), constr.get(),
1589                     enforcedRotation ? enforcedRotation->getLegacyEnfrot() : nullptr,
1590                     deform.get(),
1591                     mdModules_->outputProvider(),
1592                     mdModules_->notifier(),
1593                     inputrec, imdSession.get(), pull_work, swap, &mtop,
1594                     fcd,
1595                     globalState.get(),
1596                     &observablesHistory,
1597                     mdAtoms.get(), &nrnb, wcycle, fr,
1598                     &enerd,
1599                     &ekind,
1600                     &runScheduleWork,
1601                     replExParams,
1602                     membed,
1603                     walltime_accounting,
1604                     std::move(stopHandlerBuilder_),
1605                     doRerun,
1606                     useGpuForUpdate);
1607         simulator->run();
1608
1609         if (inputrec->bPull)
1610         {
1611             finish_pull(pull_work);
1612         }
1613         finish_swapcoords(swap);
1614     }
1615     else
1616     {
1617         GMX_RELEASE_ASSERT(pmedata, "pmedata was NULL while cr->duty was not DUTY_PP");
1618         /* do PME only */
1619         walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(emntPME));
1620         gmx_pmeonly(pmedata, cr, &nrnb, wcycle, walltime_accounting, inputrec, pmeRunMode);
1621     }
1622
1623     wallcycle_stop(wcycle, ewcRUN);
1624
1625     /* Finish up, write some stuff
1626      * if rerunMD, don't write last frame again
1627      */
1628     finish_run(fplog, mdlog, cr,
1629                inputrec, &nrnb, wcycle, walltime_accounting,
1630                fr ? fr->nbv.get() : nullptr,
1631                pmedata,
1632                EI_DYNAMICS(inputrec->eI) && !isMultiSim(ms));
1633
1634     // clean up cycle counter
1635     wallcycle_destroy(wcycle);
1636
1637 // Free PME data
1638     if (pmedata)
1639     {
1640         gmx_pme_destroy(pmedata);
1641         pmedata = nullptr;
1642     }
1643
1644     // FIXME: this is only here to manually unpin mdAtoms->chargeA_ and state->x,
1645     // before we destroy the GPU context(s) in free_gpu_resources().
1646     // Pinned buffers are associated with contexts in CUDA.
1647     // As soon as we destroy GPU contexts after mdrunner() exits, these lines should go.
1648     mdAtoms.reset(nullptr);
1649     globalState.reset(nullptr);
1650     mdModules_.reset(nullptr);   // destruct force providers here as they might also use the GPU
1651
1652     /* Free GPU memory and set a physical node tMPI barrier (which should eventually go away) */
1653     free_gpu_resources(fr, physicalNodeComm, hwinfo->gpu_info);
1654     free_gpu(nonbondedDeviceInfo);
1655     free_gpu(pmeDeviceInfo);
1656     done_forcerec(fr, mtop.molblock.size());
1657     sfree(fcd);
1658
1659     if (doMembed)
1660     {
1661         free_membed(membed);
1662     }
1663
1664     /* Does what it says */
1665     print_date_and_time(fplog, cr->nodeid, "Finished mdrun", gmx_gettime());
1666     walltime_accounting_destroy(walltime_accounting);
1667
1668     // Ensure log file content is written
1669     if (logFileHandle)
1670     {
1671         gmx_fio_flush(logFileHandle);
1672     }
1673
1674     /* Reset FPEs (important for unit tests) by disabling them. Assumes no
1675      * exceptions were enabled before function was called. */
1676     if (bEnableFPE)
1677     {
1678         gmx_fedisableexcept();
1679     }
1680
1681     auto rc = static_cast<int>(gmx_get_stop_condition());
1682
1683 #if GMX_THREAD_MPI
1684     /* we need to join all threads. The sub-threads join when they
1685        exit this function, but the master thread needs to be told to
1686        wait for that. */
1687     if (PAR(cr) && MASTER(cr))
1688     {
1689         tMPI_Finalize();
1690     }
1691 #endif
1692     return rc;
1693 }
1694
1695 Mdrunner::~Mdrunner()
1696 {
1697     // Clean up of the Manager.
1698     // This will end up getting called on every thread-MPI rank, which is unnecessary,
1699     // but okay as long as threads synchronize some time before adding or accessing
1700     // a new set of restraints.
1701     if (restraintManager_)
1702     {
1703         restraintManager_->clear();
1704         GMX_ASSERT(restraintManager_->countRestraints() == 0,
1705                    "restraints added during runner life time should be cleared at runner destruction.");
1706     }
1707 };
1708
1709 void Mdrunner::addPotential(std::shared_ptr<gmx::IRestraintPotential> puller,
1710                             const std::string                        &name)
1711 {
1712     GMX_ASSERT(restraintManager_, "Mdrunner must have a restraint manager.");
1713     // Not sure if this should be logged through the md logger or something else,
1714     // but it is helpful to have some sort of INFO level message sent somewhere.
1715     //    std::cout << "Registering restraint named " << name << std::endl;
1716
1717     // When multiple restraints are used, it may be wasteful to register them separately.
1718     // Maybe instead register an entire Restraint Manager as a force provider.
1719     restraintManager_->addToSpec(std::move(puller),
1720                                  name);
1721 }
1722
1723 Mdrunner::Mdrunner(std::unique_ptr<MDModules> mdModules)
1724     : mdModules_(std::move(mdModules))
1725 {
1726 }
1727
1728 Mdrunner::Mdrunner(Mdrunner &&) noexcept = default;
1729
1730 //NOLINTNEXTLINE(performance-noexcept-move-constructor) working around GCC bug 58265
1731 Mdrunner &Mdrunner::operator=(Mdrunner && /*handle*/) noexcept(BUGFREE_NOEXCEPT_STRING) = default;
1732
1733 class Mdrunner::BuilderImplementation
1734 {
1735     public:
1736         BuilderImplementation() = delete;
1737         BuilderImplementation(std::unique_ptr<MDModules> mdModules);
1738         ~BuilderImplementation();
1739
1740         BuilderImplementation &setExtraMdrunOptions(const MdrunOptions &options,
1741                                                     real                forceWarningThreshold,
1742                                                     StartingBehavior    startingBehavior);
1743
1744         void addDomdec(const DomdecOptions &options);
1745
1746         void addVerletList(int nstlist);
1747
1748         void addReplicaExchange(const ReplicaExchangeParameters &params);
1749
1750         void addMultiSim(gmx_multisim_t* multisim);
1751
1752         void addCommunicationRecord(t_commrec *cr);
1753
1754         void addNonBonded(const char* nbpu_opt);
1755
1756         void addPME(const char* pme_opt_, const char* pme_fft_opt_);
1757
1758         void addBondedTaskAssignment(const char* bonded_opt);
1759
1760         void addUpdateTaskAssignment(const char* update_opt);
1761
1762         void addHardwareOptions(const gmx_hw_opt_t &hardwareOptions);
1763
1764         void addFilenames(ArrayRef <const t_filenm> filenames);
1765
1766         void addOutputEnvironment(gmx_output_env_t* outputEnvironment);
1767
1768         void addLogFile(t_fileio *logFileHandle);
1769
1770         void addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder);
1771
1772         Mdrunner build();
1773
1774     private:
1775
1776         // Default parameters copied from runner.h
1777         // \todo Clarify source(s) of default parameters.
1778
1779         const char* nbpu_opt_          = nullptr;
1780         const char* pme_opt_           = nullptr;
1781         const char* pme_fft_opt_       = nullptr;
1782         const char *bonded_opt_        = nullptr;
1783         const char *update_opt_        = nullptr;
1784
1785         MdrunOptions                          mdrunOptions_;
1786
1787         DomdecOptions                         domdecOptions_;
1788
1789         ReplicaExchangeParameters             replicaExchangeParameters_;
1790
1791         //! Command-line override for the duration of a neighbor list with the Verlet scheme.
1792         int         nstlist_ = 0;
1793
1794         //! Multisim communicator handle.
1795         std::unique_ptr<gmx_multisim_t*>      multisim_ = nullptr;
1796
1797         //! Non-owning communication record (only used when building command-line mdrun)
1798         t_commrec *communicationRecord_ = nullptr;
1799
1800         //! Print a warning if any force is larger than this (in kJ/mol nm).
1801         real forceWarningThreshold_ = -1;
1802
1803         //! Whether the simulation will start afresh, or restart with/without appending.
1804         StartingBehavior startingBehavior_ = StartingBehavior::NewSimulation;
1805
1806         //! The modules that comprise the functionality of mdrun.
1807         std::unique_ptr<MDModules> mdModules_;
1808
1809         //! \brief Parallelism information.
1810         gmx_hw_opt_t hardwareOptions_;
1811
1812         //! filename options for simulation.
1813         ArrayRef<const t_filenm> filenames_;
1814
1815         /*! \brief Handle to output environment.
1816          *
1817          * \todo gmx_output_env_t needs lifetime management.
1818          */
1819         gmx_output_env_t*    outputEnvironment_ = nullptr;
1820
1821         /*! \brief Non-owning handle to MD log file.
1822          *
1823          * \todo Context should own output facilities for client.
1824          * \todo Improve log file handle management.
1825          * \internal
1826          * Code managing the FILE* relies on the ability to set it to
1827          * nullptr to check whether the filehandle is valid.
1828          */
1829         t_fileio* logFileHandle_ = nullptr;
1830
1831         /*!
1832          * \brief Builder for simulation stop signal handler.
1833          */
1834         std::unique_ptr<StopHandlerBuilder> stopHandlerBuilder_ = nullptr;
1835 };
1836
1837 Mdrunner::BuilderImplementation::BuilderImplementation(std::unique_ptr<MDModules> mdModules) :
1838     mdModules_(std::move(mdModules))
1839 {
1840 }
1841
1842 Mdrunner::BuilderImplementation::~BuilderImplementation() = default;
1843
1844 Mdrunner::BuilderImplementation &
1845 Mdrunner::BuilderImplementation::setExtraMdrunOptions(const MdrunOptions    &options,
1846                                                       const real             forceWarningThreshold,
1847                                                       const StartingBehavior startingBehavior)
1848 {
1849     mdrunOptions_          = options;
1850     forceWarningThreshold_ = forceWarningThreshold;
1851     startingBehavior_      = startingBehavior;
1852     return *this;
1853 }
1854
1855 void Mdrunner::BuilderImplementation::addDomdec(const DomdecOptions &options)
1856 {
1857     domdecOptions_ = options;
1858 }
1859
1860 void Mdrunner::BuilderImplementation::addVerletList(int nstlist)
1861 {
1862     nstlist_ = nstlist;
1863 }
1864
1865 void Mdrunner::BuilderImplementation::addReplicaExchange(const ReplicaExchangeParameters &params)
1866 {
1867     replicaExchangeParameters_ = params;
1868 }
1869
1870 void Mdrunner::BuilderImplementation::addMultiSim(gmx_multisim_t* multisim)
1871 {
1872     multisim_ = std::make_unique<gmx_multisim_t*>(multisim);
1873 }
1874
1875 void Mdrunner::BuilderImplementation::addCommunicationRecord(t_commrec *cr)
1876 {
1877     communicationRecord_ = cr;
1878 }
1879
1880 Mdrunner Mdrunner::BuilderImplementation::build()
1881 {
1882     auto newRunner = Mdrunner(std::move(mdModules_));
1883
1884     newRunner.mdrunOptions          = mdrunOptions_;
1885     newRunner.pforce                = forceWarningThreshold_;
1886     newRunner.startingBehavior      = startingBehavior_;
1887     newRunner.domdecOptions         = domdecOptions_;
1888
1889     // \todo determine an invariant to check or confirm that all gmx_hw_opt_t objects are valid
1890     newRunner.hw_opt          = hardwareOptions_;
1891
1892     // No invariant to check. This parameter exists to optionally override other behavior.
1893     newRunner.nstlist_cmdline = nstlist_;
1894
1895     newRunner.replExParams    = replicaExchangeParameters_;
1896
1897     newRunner.filenames = filenames_;
1898
1899     GMX_ASSERT(communicationRecord_, "Bug found. It should not be possible to call build() without a valid communicationRecord_.");
1900     newRunner.cr = communicationRecord_;
1901
1902     if (multisim_)
1903     {
1904         // nullptr is a valid value for the multisim handle, so we don't check the pointed-to pointer.
1905         newRunner.ms = *multisim_;
1906     }
1907     else
1908     {
1909         GMX_THROW(gmx::APIError("MdrunnerBuilder::addMultiSim() is required before build()"));
1910     }
1911
1912     // \todo Clarify ownership and lifetime management for gmx_output_env_t
1913     // \todo Update sanity checking when output environment has clearly specified invariants.
1914     // Initialization and default values for oenv are not well specified in the current version.
1915     if (outputEnvironment_)
1916     {
1917         newRunner.oenv  = outputEnvironment_;
1918     }
1919     else
1920     {
1921         GMX_THROW(gmx::APIError("MdrunnerBuilder::addOutputEnvironment() is required before build()"));
1922     }
1923
1924     newRunner.logFileHandle = logFileHandle_;
1925
1926     if (nbpu_opt_)
1927     {
1928         newRunner.nbpu_opt    = nbpu_opt_;
1929     }
1930     else
1931     {
1932         GMX_THROW(gmx::APIError("MdrunnerBuilder::addNonBonded() is required before build()"));
1933     }
1934
1935     if (pme_opt_ && pme_fft_opt_)
1936     {
1937         newRunner.pme_opt     = pme_opt_;
1938         newRunner.pme_fft_opt = pme_fft_opt_;
1939     }
1940     else
1941     {
1942         GMX_THROW(gmx::APIError("MdrunnerBuilder::addElectrostatics() is required before build()"));
1943     }
1944
1945     if (bonded_opt_)
1946     {
1947         newRunner.bonded_opt = bonded_opt_;
1948     }
1949     else
1950     {
1951         GMX_THROW(gmx::APIError("MdrunnerBuilder::addBondedTaskAssignment() is required before build()"));
1952     }
1953
1954     if (update_opt_)
1955     {
1956         newRunner.update_opt = update_opt_;
1957     }
1958     else
1959     {
1960         GMX_THROW(gmx::APIError("MdrunnerBuilder::addUpdateTaskAssignment() is required before build()  "));
1961     }
1962
1963
1964     newRunner.restraintManager_ = std::make_unique<gmx::RestraintManager>();
1965
1966     if (stopHandlerBuilder_)
1967     {
1968         newRunner.stopHandlerBuilder_ = std::move(stopHandlerBuilder_);
1969     }
1970     else
1971     {
1972         newRunner.stopHandlerBuilder_ = std::make_unique<StopHandlerBuilder>();
1973     }
1974
1975     return newRunner;
1976 }
1977
1978 void Mdrunner::BuilderImplementation::addNonBonded(const char* nbpu_opt)
1979 {
1980     nbpu_opt_ = nbpu_opt;
1981 }
1982
1983 void Mdrunner::BuilderImplementation::addPME(const char* pme_opt,
1984                                              const char* pme_fft_opt)
1985 {
1986     pme_opt_     = pme_opt;
1987     pme_fft_opt_ = pme_fft_opt;
1988 }
1989
1990 void Mdrunner::BuilderImplementation::addBondedTaskAssignment(const char* bonded_opt)
1991 {
1992     bonded_opt_ = bonded_opt;
1993 }
1994
1995 void Mdrunner::BuilderImplementation::addUpdateTaskAssignment(const char* update_opt)
1996 {
1997     update_opt_ = update_opt;
1998 }
1999
2000 void Mdrunner::BuilderImplementation::addHardwareOptions(const gmx_hw_opt_t &hardwareOptions)
2001 {
2002     hardwareOptions_ = hardwareOptions;
2003 }
2004
2005 void Mdrunner::BuilderImplementation::addFilenames(ArrayRef<const t_filenm> filenames)
2006 {
2007     filenames_ = filenames;
2008 }
2009
2010 void Mdrunner::BuilderImplementation::addOutputEnvironment(gmx_output_env_t* outputEnvironment)
2011 {
2012     outputEnvironment_ = outputEnvironment;
2013 }
2014
2015 void Mdrunner::BuilderImplementation::addLogFile(t_fileio *logFileHandle)
2016 {
2017     logFileHandle_ = logFileHandle;
2018 }
2019
2020 void Mdrunner::BuilderImplementation::addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder)
2021 {
2022     stopHandlerBuilder_ = std::move(builder);
2023 }
2024
2025 MdrunnerBuilder::MdrunnerBuilder(std::unique_ptr<MDModules>           mdModules) :
2026     impl_ {std::make_unique<Mdrunner::BuilderImplementation>(std::move(mdModules))}
2027 {
2028 }
2029
2030 MdrunnerBuilder::~MdrunnerBuilder() = default;
2031
2032 MdrunnerBuilder &MdrunnerBuilder::addSimulationMethod(const MdrunOptions    &options,
2033                                                       real                   forceWarningThreshold,
2034                                                       const StartingBehavior startingBehavior)
2035 {
2036     impl_->setExtraMdrunOptions(options, forceWarningThreshold, startingBehavior);
2037     return *this;
2038 }
2039
2040 MdrunnerBuilder &MdrunnerBuilder::addDomainDecomposition(const DomdecOptions &options)
2041 {
2042     impl_->addDomdec(options);
2043     return *this;
2044 }
2045
2046 MdrunnerBuilder &MdrunnerBuilder::addNeighborList(int nstlist)
2047 {
2048     impl_->addVerletList(nstlist);
2049     return *this;
2050 }
2051
2052 MdrunnerBuilder &MdrunnerBuilder::addReplicaExchange(const ReplicaExchangeParameters &params)
2053 {
2054     impl_->addReplicaExchange(params);
2055     return *this;
2056 }
2057
2058 MdrunnerBuilder &MdrunnerBuilder::addMultiSim(gmx_multisim_t* multisim)
2059 {
2060     impl_->addMultiSim(multisim);
2061     return *this;
2062 }
2063
2064 MdrunnerBuilder &MdrunnerBuilder::addCommunicationRecord(t_commrec *cr)
2065 {
2066     impl_->addCommunicationRecord(cr);
2067     return *this;
2068 }
2069
2070 MdrunnerBuilder &MdrunnerBuilder::addNonBonded(const char* nbpu_opt)
2071 {
2072     impl_->addNonBonded(nbpu_opt);
2073     return *this;
2074 }
2075
2076 MdrunnerBuilder &MdrunnerBuilder::addElectrostatics(const char* pme_opt,
2077                                                     const char* pme_fft_opt)
2078 {
2079     // The builder method may become more general in the future, but in this version,
2080     // parameters for PME electrostatics are both required and the only parameters
2081     // available.
2082     if (pme_opt && pme_fft_opt)
2083     {
2084         impl_->addPME(pme_opt, pme_fft_opt);
2085     }
2086     else
2087     {
2088         GMX_THROW(gmx::InvalidInputError("addElectrostatics() arguments must be non-null pointers."));
2089     }
2090     return *this;
2091 }
2092
2093 MdrunnerBuilder &MdrunnerBuilder::addBondedTaskAssignment(const char* bonded_opt)
2094 {
2095     impl_->addBondedTaskAssignment(bonded_opt);
2096     return *this;
2097 }
2098
2099 MdrunnerBuilder &MdrunnerBuilder::addUpdateTaskAssignment(const char* update_opt)
2100 {
2101     impl_->addUpdateTaskAssignment(update_opt);
2102     return *this;
2103 }
2104
2105 Mdrunner MdrunnerBuilder::build()
2106 {
2107     return impl_->build();
2108 }
2109
2110 MdrunnerBuilder &MdrunnerBuilder::addHardwareOptions(const gmx_hw_opt_t &hardwareOptions)
2111 {
2112     impl_->addHardwareOptions(hardwareOptions);
2113     return *this;
2114 }
2115
2116 MdrunnerBuilder &MdrunnerBuilder::addFilenames(ArrayRef<const t_filenm> filenames)
2117 {
2118     impl_->addFilenames(filenames);
2119     return *this;
2120 }
2121
2122 MdrunnerBuilder &MdrunnerBuilder::addOutputEnvironment(gmx_output_env_t* outputEnvironment)
2123 {
2124     impl_->addOutputEnvironment(outputEnvironment);
2125     return *this;
2126 }
2127
2128 MdrunnerBuilder &MdrunnerBuilder::addLogFile(t_fileio *logFileHandle)
2129 {
2130     impl_->addLogFile(logFileHandle);
2131     return *this;
2132 }
2133
2134 MdrunnerBuilder &MdrunnerBuilder::addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder)
2135 {
2136     impl_->addStopHandlerBuilder(std::move(builder));
2137     return *this;
2138 }
2139
2140 MdrunnerBuilder::MdrunnerBuilder(MdrunnerBuilder &&) noexcept = default;
2141
2142 MdrunnerBuilder &MdrunnerBuilder::operator=(MdrunnerBuilder &&) noexcept = default;
2143
2144 } // namespace gmx