src/gromacs/mdlib/sim_util.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) 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 #include "gmxpre.h"
  38
  39 #include "config.h"
  40
  41 #include <cmath>
  42 #include <cstdint>
  43 #include <cstdio>
  44 #include <cstring>
  45
  46 #include <array>
  47
  48 #include "gromacs/awh/awh.h"
  49 #include "gromacs/domdec/dlbtiming.h"
  50 #include "gromacs/domdec/domdec.h"
  51 #include "gromacs/domdec/domdec_struct.h"
  52 #include "gromacs/domdec/partition.h"
  53 #include "gromacs/essentialdynamics/edsam.h"
  54 #include "gromacs/ewald/pme.h"
  55 #include "gromacs/gmxlib/chargegroup.h"
  56 #include "gromacs/gmxlib/network.h"
  57 #include "gromacs/gmxlib/nrnb.h"
  58 #include "gromacs/gmxlib/nonbonded/nb_free_energy.h"
  59 #include "gromacs/gmxlib/nonbonded/nb_kernel.h"
  60 #include "gromacs/gmxlib/nonbonded/nonbonded.h"
  61 #include "gromacs/gpu_utils/gpu_utils.h"
  62 #include "gromacs/imd/imd.h"
  63 #include "gromacs/listed_forces/bonded.h"
  64 #include "gromacs/listed_forces/disre.h"
  65 #include "gromacs/listed_forces/gpubonded.h"
  66 #include "gromacs/listed_forces/listed_forces.h"
  67 #include "gromacs/listed_forces/manage_threading.h"
  68 #include "gromacs/listed_forces/orires.h"
  69 #include "gromacs/math/arrayrefwithpadding.h"
  70 #include "gromacs/math/functions.h"
  71 #include "gromacs/math/units.h"
  72 #include "gromacs/math/vec.h"
  73 #include "gromacs/math/vecdump.h"
  74 #include "gromacs/mdlib/calcmu.h"
  75 #include "gromacs/mdlib/calcvir.h"
  76 #include "gromacs/mdlib/constr.h"
  77 #include "gromacs/mdlib/enerdata_utils.h"
  78 #include "gromacs/mdlib/force.h"
  79 #include "gromacs/mdlib/forcerec.h"
  80 #include "gromacs/mdlib/gmx_omp_nthreads.h"
  81 #include "gromacs/mdlib/ppforceworkload.h"
  82 #include "gromacs/mdlib/qmmm.h"
  83 #include "gromacs/mdlib/update.h"
  84 #include "gromacs/mdtypes/commrec.h"
  85 #include "gromacs/mdtypes/enerdata.h"
  86 #include "gromacs/mdtypes/forceoutput.h"
  87 #include "gromacs/mdtypes/iforceprovider.h"
  88 #include "gromacs/mdtypes/inputrec.h"
  89 #include "gromacs/mdtypes/md_enums.h"
  90 #include "gromacs/mdtypes/state.h"
  91 #include "gromacs/nbnxm/atomdata.h"
  92 #include "gromacs/nbnxm/gpu_data_mgmt.h"
  93 #include "gromacs/nbnxm/nbnxm.h"
  94 #include "gromacs/pbcutil/ishift.h"
  95 #include "gromacs/pbcutil/mshift.h"
  96 #include "gromacs/pbcutil/pbc.h"
  97 #include "gromacs/pulling/pull.h"
  98 #include "gromacs/pulling/pull_rotation.h"
  99 #include "gromacs/timing/cyclecounter.h"
 100 #include "gromacs/timing/gpu_timing.h"
 101 #include "gromacs/timing/wallcycle.h"
 102 #include "gromacs/timing/wallcyclereporting.h"
 103 #include "gromacs/timing/walltime_accounting.h"
 104 #include "gromacs/topology/topology.h"
 105 #include "gromacs/utility/arrayref.h"
 106 #include "gromacs/utility/basedefinitions.h"
 107 #include "gromacs/utility/cstringutil.h"
 108 #include "gromacs/utility/exceptions.h"
 109 #include "gromacs/utility/fatalerror.h"
 110 #include "gromacs/utility/gmxassert.h"
 111 #include "gromacs/utility/gmxmpi.h"
 112 #include "gromacs/utility/logger.h"
 113 #include "gromacs/utility/smalloc.h"
 114 #include "gromacs/utility/strconvert.h"
 115 #include "gromacs/utility/sysinfo.h"
 116
 117 // TODO: this environment variable allows us to verify before release
 118 // that on less common architectures the total cost of polling is not larger than
 119 // a blocking wait (so polling does not introduce overhead when the static
 120 // PME-first ordering would suffice).
 121 static const bool c_disableAlternatingWait = (getenv("GMX_DISABLE_ALTERNATING_GPU_WAIT") != nullptr);
 122
 123 // environment variable to enable GPU buffer ops, to allow incremental and optional
 124 // introduction of this functionality.
 125 // TODO eventially tie this in with other existing GPU flags.
 126 static const bool c_enableGpuBufOps = (getenv("GMX_USE_GPU_BUFFER_OPS") != nullptr);
 127
 128 static void sum_forces(rvec f[], gmx::ArrayRef<const gmx::RVec> forceToAdd)
 129 {
 130     const int      end = forceToAdd.size();
 131
 132     int gmx_unused nt = gmx_omp_nthreads_get(emntDefault);
 133 #pragma omp parallel for num_threads(nt) schedule(static)
 134     for (int i = 0; i < end; i++)
 135     {
 136         rvec_inc(f[i], forceToAdd[i]);
 137     }
 138 }
 139
 140 static void calc_virial(int start, int homenr, const rvec x[], const rvec f[],
 141                         tensor vir_part, const t_graph *graph, const matrix box,
 142                         t_nrnb *nrnb, const t_forcerec *fr, int ePBC)
 143 {
 144     /* The short-range virial from surrounding boxes */
 145     calc_vir(SHIFTS, fr->shift_vec, fr->fshift, vir_part, ePBC == epbcSCREW, box);
 146     inc_nrnb(nrnb, eNR_VIRIAL, SHIFTS);
 147
 148     /* Calculate partial virial, for local atoms only, based on short range.
 149      * Total virial is computed in global_stat, called from do_md
 150      */
 151     f_calc_vir(start, start+homenr, x, f, vir_part, graph, box);
 152     inc_nrnb(nrnb, eNR_VIRIAL, homenr);
 153
 154     if (debug)
 155     {
 156         pr_rvecs(debug, 0, "vir_part", vir_part, DIM);
 157     }
 158 }
 159
 160 static void pull_potential_wrapper(const t_commrec *cr,
 161                                    const t_inputrec *ir,
 162                                    const matrix box, gmx::ArrayRef<const gmx::RVec> x,
 163                                    gmx::ForceWithVirial *force,
 164                                    const t_mdatoms *mdatoms,
 165                                    gmx_enerdata_t *enerd,
 166                                    pull_t *pull_work,
 167                                    const real *lambda,
 168                                    double t,
 169                                    gmx_wallcycle_t wcycle)
 170 {
 171     t_pbc  pbc;
 172     real   dvdl;
 173
 174     /* Calculate the center of mass forces, this requires communication,
 175      * which is why pull_potential is called close to other communication.
 176      */
 177     wallcycle_start(wcycle, ewcPULLPOT);
 178     set_pbc(&pbc, ir->ePBC, box);
 179     dvdl                     = 0;
 180     enerd->term[F_COM_PULL] +=
 181         pull_potential(pull_work, mdatoms, &pbc,
 182                        cr, t, lambda[efptRESTRAINT], as_rvec_array(x.data()), force, &dvdl);
 183     enerd->dvdl_lin[efptRESTRAINT] += dvdl;
 184     wallcycle_stop(wcycle, ewcPULLPOT);
 185 }
 186
 187 static void pme_receive_force_ener(const t_commrec      *cr,
 188                                    gmx::ForceWithVirial *forceWithVirial,
 189                                    gmx_enerdata_t       *enerd,
 190                                    gmx_wallcycle_t       wcycle)
 191 {
 192     real   e_q, e_lj, dvdl_q, dvdl_lj;
 193     float  cycles_ppdpme, cycles_seppme;
 194
 195     cycles_ppdpme = wallcycle_stop(wcycle, ewcPPDURINGPME);
 196     dd_cycles_add(cr->dd, cycles_ppdpme, ddCyclPPduringPME);
 197
 198     /* In case of node-splitting, the PP nodes receive the long-range
 199      * forces, virial and energy from the PME nodes here.
 200      */
 201     wallcycle_start(wcycle, ewcPP_PMEWAITRECVF);
 202     dvdl_q  = 0;
 203     dvdl_lj = 0;
 204     gmx_pme_receive_f(cr, forceWithVirial, &e_q, &e_lj, &dvdl_q, &dvdl_lj,
 205                       &cycles_seppme);
 206     enerd->term[F_COUL_RECIP] += e_q;
 207     enerd->term[F_LJ_RECIP]   += e_lj;
 208     enerd->dvdl_lin[efptCOUL] += dvdl_q;
 209     enerd->dvdl_lin[efptVDW]  += dvdl_lj;
 210
 211     if (wcycle)
 212     {
 213         dd_cycles_add(cr->dd, cycles_seppme, ddCyclPME);
 214     }
 215     wallcycle_stop(wcycle, ewcPP_PMEWAITRECVF);
 216 }
 217
 218 static void print_large_forces(FILE            *fp,
 219                                const t_mdatoms *md,
 220                                const t_commrec *cr,
 221                                int64_t          step,
 222                                real             forceTolerance,
 223                                const rvec      *x,
 224                                const rvec      *f)
 225 {
 226     real           force2Tolerance = gmx::square(forceTolerance);
 227     gmx::index     numNonFinite    = 0;
 228     for (int i = 0; i < md->homenr; i++)
 229     {
 230         real force2    = norm2(f[i]);
 231         bool nonFinite = !std::isfinite(force2);
 232         if (force2 >= force2Tolerance || nonFinite)
 233         {
 234             fprintf(fp, "step %" PRId64 " atom %6d  x %8.3f %8.3f %8.3f  force %12.5e\n",
 235                     step,
 236                     ddglatnr(cr->dd, i), x[i][XX], x[i][YY], x[i][ZZ], std::sqrt(force2));
 237         }
 238         if (nonFinite)
 239         {
 240             numNonFinite++;
 241         }
 242     }
 243     if (numNonFinite > 0)
 244     {
 245         /* Note that with MPI this fatal call on one rank might interrupt
 246          * the printing on other ranks. But we can only avoid that with
 247          * an expensive MPI barrier that we would need at each step.
 248          */
 249         gmx_fatal(FARGS, "At step %" PRId64 " detected non-finite forces on %td atoms", step, numNonFinite);
 250     }
 251 }
 252
 253 static void post_process_forces(const t_commrec           *cr,
 254                                 int64_t                    step,
 255                                 t_nrnb                    *nrnb,
 256                                 gmx_wallcycle_t            wcycle,
 257                                 const gmx_localtop_t      *top,
 258                                 const matrix               box,
 259                                 const rvec                 x[],
 260                                 rvec                       f[],
 261                                 gmx::ForceWithVirial      *forceWithVirial,
 262                                 tensor                     vir_force,
 263                                 const t_mdatoms           *mdatoms,
 264                                 const t_graph             *graph,
 265                                 const t_forcerec          *fr,
 266                                 const gmx_vsite_t         *vsite,
 267                                 int                        flags)
 268 {
 269     if (fr->haveDirectVirialContributions)
 270     {
 271         rvec *fDirectVir = as_rvec_array(forceWithVirial->force_.data());
 272
 273         if (vsite)
 274         {
 275             /* Spread the mesh force on virtual sites to the other particles...
 276              * This is parallellized. MPI communication is performed
 277              * if the constructing atoms aren't local.
 278              */
 279             matrix virial = { { 0 } };
 280             spread_vsite_f(vsite, x, fDirectVir, nullptr,
 281                            (flags & GMX_FORCE_VIRIAL) != 0, virial,
 282                            nrnb,
 283                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr, wcycle);
 284             forceWithVirial->addVirialContribution(virial);
 285         }
 286
 287         if (flags & GMX_FORCE_VIRIAL)
 288         {
 289             /* Now add the forces, this is local */
 290             sum_forces(f, forceWithVirial->force_);
 291
 292             /* Add the direct virial contributions */
 293             GMX_ASSERT(forceWithVirial->computeVirial_, "forceWithVirial should request virial computation when we request the virial");
 294             m_add(vir_force, forceWithVirial->getVirial(), vir_force);
 295
 296             if (debug)
 297             {
 298                 pr_rvecs(debug, 0, "vir_force", vir_force, DIM);
 299             }
 300         }
 301     }
 302
 303     if (fr->print_force >= 0)
 304     {
 305         print_large_forces(stderr, mdatoms, cr, step, fr->print_force, x, f);
 306     }
 307 }
 308
 309 static void do_nb_verlet(t_forcerec                       *fr,
 310                          const interaction_const_t        *ic,
 311                          gmx_enerdata_t                   *enerd,
 312                          const int                         flags,
 313                          const Nbnxm::InteractionLocality  ilocality,
 314                          const int                         clearF,
 315                          const int64_t                     step,
 316                          t_nrnb                           *nrnb,
 317                          gmx_wallcycle_t                   wcycle)
 318 {
 319     if (!(flags & GMX_FORCE_NONBONDED))
 320     {
 321         /* skip non-bonded calculation */
 322         return;
 323     }
 324
 325     nonbonded_verlet_t *nbv  = fr->nbv.get();
 326
 327     /* GPU kernel launch overhead is already timed separately */
 328     if (fr->cutoff_scheme != ecutsVERLET)
 329     {
 330         gmx_incons("Invalid cut-off scheme passed!");
 331     }
 332
 333     if (!nbv->useGpu())
 334     {
 335         /* When dynamic pair-list  pruning is requested, we need to prune
 336          * at nstlistPrune steps.
 337          */
 338         if (nbv->isDynamicPruningStepCpu(step))
 339         {
 340             /* Prune the pair-list beyond fr->ic->rlistPrune using
 341              * the current coordinates of the atoms.
 342              */
 343             wallcycle_sub_start(wcycle, ewcsNONBONDED_PRUNING);
 344             nbv->dispatchPruneKernelCpu(ilocality, fr->shift_vec);
 345             wallcycle_sub_stop(wcycle, ewcsNONBONDED_PRUNING);
 346         }
 347     }
 348
 349     nbv->dispatchNonbondedKernel(ilocality, *ic, flags, clearF, *fr, enerd, nrnb, wcycle);
 350 }
 351
 352 static inline void clear_rvecs_omp(int n, rvec v[])
 353 {
 354     int nth = gmx_omp_nthreads_get_simple_rvec_task(emntDefault, n);
 355
 356     /* Note that we would like to avoid this conditional by putting it
 357      * into the omp pragma instead, but then we still take the full
 358      * omp parallel for overhead (at least with gcc5).
 359      */
 360     if (nth == 1)
 361     {
 362         for (int i = 0; i < n; i++)
 363         {
 364             clear_rvec(v[i]);
 365         }
 366     }
 367     else
 368     {
 369 #pragma omp parallel for num_threads(nth) schedule(static)
 370         for (int i = 0; i < n; i++)
 371         {
 372             clear_rvec(v[i]);
 373         }
 374     }
 375 }
 376
 377 /*! \brief Return an estimate of the average kinetic energy or 0 when unreliable
 378  *
 379  * \param groupOptions  Group options, containing T-coupling options
 380  */
 381 static real averageKineticEnergyEstimate(const t_grpopts &groupOptions)
 382 {
 383     real nrdfCoupled   = 0;
 384     real nrdfUncoupled = 0;
 385     real kineticEnergy = 0;
 386     for (int g = 0; g < groupOptions.ngtc; g++)
 387     {
 388         if (groupOptions.tau_t[g] >= 0)
 389         {
 390             nrdfCoupled   += groupOptions.nrdf[g];
 391             kineticEnergy += groupOptions.nrdf[g]*0.5*groupOptions.ref_t[g]*BOLTZ;
 392         }
 393         else
 394         {
 395             nrdfUncoupled += groupOptions.nrdf[g];
 396         }
 397     }
 398
 399     /* This conditional with > also catches nrdf=0 */
 400     if (nrdfCoupled > nrdfUncoupled)
 401     {
 402         return kineticEnergy*(nrdfCoupled + nrdfUncoupled)/nrdfCoupled;
 403     }
 404     else
 405     {
 406         return 0;
 407     }
 408 }
 409
 410 /*! \brief This routine checks that the potential energy is finite.
 411  *
 412  * Always checks that the potential energy is finite. If step equals
 413  * inputrec.init_step also checks that the magnitude of the potential energy
 414  * is reasonable. Terminates with a fatal error when a check fails.
 415  * Note that passing this check does not guarantee finite forces,
 416  * since those use slightly different arithmetics. But in most cases
 417  * there is just a narrow coordinate range where forces are not finite
 418  * and energies are finite.
 419  *
 420  * \param[in] step      The step number, used for checking and printing
 421  * \param[in] enerd     The energy data; the non-bonded group energies need to be added to enerd.term[F_EPOT] before calling this routine
 422  * \param[in] inputrec  The input record
 423  */
 424 static void checkPotentialEnergyValidity(int64_t               step,
 425                                          const gmx_enerdata_t &enerd,
 426                                          const t_inputrec     &inputrec)
 427 {
 428     /* Threshold valid for comparing absolute potential energy against
 429      * the kinetic energy. Normally one should not consider absolute
 430      * potential energy values, but with a factor of one million
 431      * we should never get false positives.
 432      */
 433     constexpr real c_thresholdFactor = 1e6;
 434
 435     bool           energyIsNotFinite    = !std::isfinite(enerd.term[F_EPOT]);
 436     real           averageKineticEnergy = 0;
 437     /* We only check for large potential energy at the initial step,
 438      * because that is by far the most likely step for this too occur
 439      * and because computing the average kinetic energy is not free.
 440      * Note: nstcalcenergy >> 1 often does not allow to catch large energies
 441      * before they become NaN.
 442      */
 443     if (step == inputrec.init_step && EI_DYNAMICS(inputrec.eI))
 444     {
 445         averageKineticEnergy = averageKineticEnergyEstimate(inputrec.opts);
 446     }
 447
 448     if (energyIsNotFinite || (averageKineticEnergy > 0 &&
 449                               enerd.term[F_EPOT] > c_thresholdFactor*averageKineticEnergy))
 450     {
 451         gmx_fatal(FARGS, "Step %" PRId64 ": The total potential energy is %g, which is %s. The LJ and electrostatic contributions to the energy are %g and %g, respectively. A %s potential energy can be caused by overlapping interactions in bonded interactions or very large%s coordinate values. Usually this is caused by a badly- or non-equilibrated initial configuration, incorrect interactions or parameters in the topology.",
 452                   step,
 453                   enerd.term[F_EPOT],
 454                   energyIsNotFinite ? "not finite" : "extremely high",
 455                   enerd.term[F_LJ],
 456                   enerd.term[F_COUL_SR],
 457                   energyIsNotFinite ? "non-finite" : "very high",
 458                   energyIsNotFinite ? " or Nan" : "");
 459     }
 460 }
 461
 462 /*! \brief Return true if there are special forces computed this step.
 463  *
 464  * The conditionals exactly correspond to those in computeSpecialForces().
 465  */
 466 static bool
 467 haveSpecialForces(const t_inputrec              *inputrec,
 468                   ForceProviders                *forceProviders,
 469                   const pull_t                  *pull_work,
 470                   int                            forceFlags,
 471                   const gmx_edsam               *ed)
 472 {
 473     const bool computeForces = (forceFlags & GMX_FORCE_FORCES) != 0;
 474
 475     return
 476         ((computeForces && forceProviders->hasForceProvider()) ||         // forceProviders
 477          (inputrec->bPull && pull_have_potential(pull_work)) ||           // pull
 478          inputrec->bRot ||                                                // enforced rotation
 479          (ed != nullptr) ||                                               // flooding
 480          (inputrec->bIMD && computeForces));                              // IMD
 481 }
 482
 483 /*! \brief Compute forces and/or energies for special algorithms
 484  *
 485  * The intention is to collect all calls to algorithms that compute
 486  * forces on local atoms only and that do not contribute to the local
 487  * virial sum (but add their virial contribution separately).
 488  * Eventually these should likely all become ForceProviders.
 489  * Within this function the intention is to have algorithms that do
 490  * global communication at the end, so global barriers within the MD loop
 491  * are as close together as possible.
 492  *
 493  * \param[in]     fplog            The log file
 494  * \param[in]     cr               The communication record
 495  * \param[in]     inputrec         The input record
 496  * \param[in]     awh              The Awh module (nullptr if none in use).
 497  * \param[in]     enforcedRotation Enforced rotation module.
 498  * \param[in]     imdSession       The IMD session
 499  * \param[in]     pull_work        The pull work structure.
 500  * \param[in]     step             The current MD step
 501  * \param[in]     t                The current time
 502  * \param[in,out] wcycle           Wallcycle accounting struct
 503  * \param[in,out] forceProviders   Pointer to a list of force providers
 504  * \param[in]     box              The unit cell
 505  * \param[in]     x                The coordinates
 506  * \param[in]     mdatoms          Per atom properties
 507  * \param[in]     lambda           Array of free-energy lambda values
 508  * \param[in]     forceFlags       Flags that tell whether we should compute forces/energies/virial
 509  * \param[in,out] forceWithVirial  Force and virial buffers
 510  * \param[in,out] enerd            Energy buffer
 511  * \param[in,out] ed               Essential dynamics pointer
 512  * \param[in]     bNS              Tells if we did neighbor searching this step, used for ED sampling
 513  *
 514  * \todo Remove bNS, which is used incorrectly.
 515  * \todo Convert all other algorithms called here to ForceProviders.
 516  */
 517 static void
 518 computeSpecialForces(FILE                          *fplog,
 519                      const t_commrec               *cr,
 520                      const t_inputrec              *inputrec,
 521                      gmx::Awh                      *awh,
 522                      gmx_enfrot                    *enforcedRotation,
 523                      gmx::ImdSession               *imdSession,
 524                      pull_t                        *pull_work,
 525                      int64_t                        step,
 526                      double                         t,
 527                      gmx_wallcycle_t                wcycle,
 528                      ForceProviders                *forceProviders,
 529                      matrix                         box,
 530                      gmx::ArrayRef<const gmx::RVec> x,
 531                      const t_mdatoms               *mdatoms,
 532                      real                          *lambda,
 533                      int                            forceFlags,
 534                      gmx::ForceWithVirial          *forceWithVirial,
 535                      gmx_enerdata_t                *enerd,
 536                      gmx_edsam                     *ed,
 537                      gmx_bool                       bNS)
 538 {
 539     const bool computeForces = (forceFlags & GMX_FORCE_FORCES) != 0;
 540
 541     /* NOTE: Currently all ForceProviders only provide forces.
 542      *       When they also provide energies, remove this conditional.
 543      */
 544     if (computeForces)
 545     {
 546         gmx::ForceProviderInput  forceProviderInput(x, *mdatoms, t, box, *cr);
 547         gmx::ForceProviderOutput forceProviderOutput(forceWithVirial, enerd);
 548
 549         /* Collect forces from modules */
 550         forceProviders->calculateForces(forceProviderInput, &forceProviderOutput);
 551     }
 552
 553     if (inputrec->bPull && pull_have_potential(pull_work))
 554     {
 555         pull_potential_wrapper(cr, inputrec, box, x,
 556                                forceWithVirial,
 557                                mdatoms, enerd, pull_work, lambda, t,
 558                                wcycle);
 559
 560         if (awh)
 561         {
 562             enerd->term[F_COM_PULL] +=
 563                 awh->applyBiasForcesAndUpdateBias(inputrec->ePBC, *mdatoms, box,
 564                                                   forceWithVirial,
 565                                                   t, step, wcycle, fplog);
 566         }
 567     }
 568
 569     rvec *f = as_rvec_array(forceWithVirial->force_.data());
 570
 571     /* Add the forces from enforced rotation potentials (if any) */
 572     if (inputrec->bRot)
 573     {
 574         wallcycle_start(wcycle, ewcROTadd);
 575         enerd->term[F_COM_PULL] += add_rot_forces(enforcedRotation, f, cr, step, t);
 576         wallcycle_stop(wcycle, ewcROTadd);
 577     }
 578
 579     if (ed)
 580     {
 581         /* Note that since init_edsam() is called after the initialization
 582          * of forcerec, edsam doesn't request the noVirSum force buffer.
 583          * Thus if no other algorithm (e.g. PME) requires it, the forces
 584          * here will contribute to the virial.
 585          */
 586         do_flood(cr, inputrec, as_rvec_array(x.data()), f, ed, box, step, bNS);
 587     }
 588
 589     /* Add forces from interactive molecular dynamics (IMD), if any */
 590     if (inputrec->bIMD && computeForces)
 591     {
 592         imdSession->applyForces(f);
 593     }
 594 }
 595
 596 /*! \brief Launch the prepare_step and spread stages of PME GPU.
 597  *
 598  * \param[in]  pmedata       The PME structure
 599  * \param[in]  box           The box matrix
 600  * \param[in]  x             Coordinate array
 601  * \param[in]  flags         Force flags
 602  * \param[in]  pmeFlags      PME flags
 603  * \param[in]  wcycle        The wallcycle structure
 604  */
 605 static inline void launchPmeGpuSpread(gmx_pme_t      *pmedata,
 606                                       matrix          box,
 607                                       rvec            x[],
 608                                       int             flags,
 609                                       int             pmeFlags,
 610                                       gmx_wallcycle_t wcycle)
 611 {
 612     pme_gpu_prepare_computation(pmedata, (flags & GMX_FORCE_DYNAMICBOX) != 0, box, wcycle, pmeFlags);
 613     pme_gpu_launch_spread(pmedata, x, wcycle);
 614 }
 615
 616 /*! \brief Launch the FFT and gather stages of PME GPU
 617  *
 618  * This function only implements setting the output forces (no accumulation).
 619  *
 620  * \param[in]  pmedata        The PME structure
 621  * \param[in]  wcycle         The wallcycle structure
 622  */
 623 static void launchPmeGpuFftAndGather(gmx_pme_t        *pmedata,
 624                                      gmx_wallcycle_t   wcycle)
 625 {
 626     pme_gpu_launch_complex_transforms(pmedata, wcycle);
 627     pme_gpu_launch_gather(pmedata, wcycle, PmeForceOutputHandling::Set);
 628 }
 629
 630 /*! \brief
 631  *  Polling wait for either of the PME or nonbonded GPU tasks.
 632  *
 633  * Instead of a static order in waiting for GPU tasks, this function
 634  * polls checking which of the two tasks completes first, and does the
 635  * associated force buffer reduction overlapped with the other task.
 636  * By doing that, unlike static scheduling order, it can always overlap
 637  * one of the reductions, regardless of the GPU task completion order.
 638  *
 639  * \param[in]     nbv              Nonbonded verlet structure
 640  * \param[in,out] pmedata          PME module data
 641  * \param[in,out] force            Force array to reduce task outputs into.
 642  * \param[in,out] forceWithVirial  Force and virial buffers
 643  * \param[in,out] fshift           Shift force output vector results are reduced into
 644  * \param[in,out] enerd            Energy data structure results are reduced into
 645  * \param[in]     flags            Force flags
 646  * \param[in]     pmeFlags         PME flags
 647  * \param[in]     wcycle           The wallcycle structure
 648  */
 649 static void alternatePmeNbGpuWaitReduce(nonbonded_verlet_t                  *nbv,
 650                                         gmx_pme_t                           *pmedata,
 651                                         gmx::ArrayRefWithPadding<gmx::RVec> *force,
 652                                         gmx::ForceWithVirial                *forceWithVirial,
 653                                         rvec                                 fshift[],
 654                                         gmx_enerdata_t                      *enerd,
 655                                         int                                  flags,
 656                                         int                                  pmeFlags,
 657                                         gmx_wallcycle_t                      wcycle)
 658 {
 659     bool isPmeGpuDone = false;
 660     bool isNbGpuDone  = false;
 661
 662
 663     gmx::ArrayRef<const gmx::RVec> pmeGpuForces;
 664
 665     while (!isPmeGpuDone || !isNbGpuDone)
 666     {
 667         if (!isPmeGpuDone)
 668         {
 669             GpuTaskCompletion completionType = (isNbGpuDone) ? GpuTaskCompletion::Wait : GpuTaskCompletion::Check;
 670             isPmeGpuDone = pme_gpu_try_finish_task(pmedata, pmeFlags, wcycle, forceWithVirial, enerd, completionType);
 671         }
 672
 673         if (!isNbGpuDone)
 674         {
 675             GpuTaskCompletion completionType = (isPmeGpuDone) ? GpuTaskCompletion::Wait : GpuTaskCompletion::Check;
 676             wallcycle_start_nocount(wcycle, ewcWAIT_GPU_NB_L);
 677             isNbGpuDone = Nbnxm::gpu_try_finish_task(nbv->gpu_nbv,
 678                                                      flags,
 679                                                      Nbnxm::AtomLocality::Local,
 680                                                      enerd->grpp.ener[egLJSR].data(),
 681                                                      enerd->grpp.ener[egCOULSR].data(),
 682                                                      fshift, completionType);
 683             wallcycle_stop(wcycle, ewcWAIT_GPU_NB_L);
 684             // To get the call count right, when the task finished we
 685             // issue a start/stop.
 686             // TODO: move the ewcWAIT_GPU_NB_L cycle counting into nbnxn_gpu_try_finish_task()
 687             // and ewcNB_XF_BUF_OPS counting into nbnxn_atomdata_add_nbat_f_to_f().
 688             if (isNbGpuDone)
 689             {
 690                 wallcycle_start(wcycle, ewcWAIT_GPU_NB_L);
 691                 wallcycle_stop(wcycle, ewcWAIT_GPU_NB_L);
 692
 693                 nbv->atomdata_add_nbat_f_to_f(Nbnxm::AtomLocality::Local,
 694                                               as_rvec_array(force->unpaddedArrayRef().data()),
 695                                               BufferOpsUseGpu::False,
 696                                               GpuBufferOpsAccumulateForce::Null,
 697                                               wcycle);
 698             }
 699         }
 700     }
 701 }
 702
 703 /*! \brief Hack structure with force ouput buffers for do_force for the home atoms for this domain */
 704 struct ForceOutputs
 705 {
 706     //! Constructor
 707     ForceOutputs(rvec *f, gmx::ForceWithVirial const forceWithVirial) :
 708         f(f),
 709         forceWithVirial(forceWithVirial) {}
 710
 711     //! Force output buffer used by legacy modules (without SIMD padding)
 712     rvec                 *const f;
 713     //! Force with direct virial contribution (if there are any; without SIMD padding)
 714     gmx::ForceWithVirial        forceWithVirial;
 715 };
 716
 717 /*! \brief Set up the different force buffers; also does clearing.
 718  *
 719  * \param[in] fr        force record pointer
 720  * \param[in] pull_work The pull work object.
 721  * \param[in] inputrec  input record
 722  * \param[in] force     force array
 723  * \param[in] bDoForces True if force are computed this step
 724  * \param[in] doVirial  True if virial is computed this step
 725  * \param[out] wcycle   wallcycle recording structure
 726  *
 727  * \returns             Cleared force output structure
 728  */
 729 static ForceOutputs
 730 setupForceOutputs(const t_forcerec                    *fr,
 731                   pull_t                              *pull_work,
 732                   const t_inputrec                    &inputrec,
 733                   gmx::ArrayRefWithPadding<gmx::RVec>  force,
 734                   const bool                           bDoForces,
 735                   const bool                           doVirial,
 736                   gmx_wallcycle_t                      wcycle)
 737 {
 738     wallcycle_sub_start(wcycle, ewcsCLEAR_FORCE_BUFFER);
 739
 740     /* Temporary solution until all routines take PaddedRVecVector */
 741     rvec *const f = as_rvec_array(force.unpaddedArrayRef().data());
 742     if (bDoForces)
 743     {
 744         /* Clear the short- and long-range forces */
 745         clear_rvecs_omp(fr->natoms_force_constr, f);
 746     }
 747
 748     /* If we need to compute the virial, we might need a separate
 749      * force buffer for algorithms for which the virial is calculated
 750      * directly, such as PME. Otherwise, forceWithVirial uses the
 751      * the same force (f in legacy calls) buffer as other algorithms.
 752      */
 753     const bool useSeparateForceWithVirialBuffer = (bDoForces && (doVirial && fr->haveDirectVirialContributions));
 754
 755
 756     /* forceWithVirial uses the local atom range only */
 757     gmx::ForceWithVirial forceWithVirial (useSeparateForceWithVirialBuffer ?
 758                                           *fr->forceBufferForDirectVirialContributions : force.unpaddedArrayRef(),
 759                                           doVirial);
 760
 761     if (useSeparateForceWithVirialBuffer)
 762     {
 763         /* TODO: update comment
 764          * We only compute forces on local atoms. Note that vsites can
 765          * spread to non-local atoms, but that part of the buffer is
 766          * cleared separately in the vsite spreading code.
 767          */
 768         clear_rvecs_omp(forceWithVirial.force_.size(), as_rvec_array(forceWithVirial.force_.data()));
 769     }
 770
 771     if (inputrec.bPull && pull_have_constraint(pull_work))
 772     {
 773         clear_pull_forces(pull_work);
 774     }
 775
 776     wallcycle_sub_stop(wcycle, ewcsCLEAR_FORCE_BUFFER);
 777
 778     return ForceOutputs(f, forceWithVirial);
 779 }
 780
 781
 782 /*! \brief Set up flags that indicate what type of work is there to compute.
 783  *
 784  * Currently we only update it at search steps,
 785  * but some properties may change more frequently (e.g. virial/non-virial step),
 786  * so when including those either the frequency of update (per-step) or the scope
 787  * of a flag will change (i.e. a set of flags for nstlist steps).
 788  *
 789  */
 790 static void
 791 setupForceWorkload(gmx::PpForceWorkload *forceWork,
 792                    const t_inputrec     *inputrec,
 793                    const t_forcerec     *fr,
 794                    const pull_t         *pull_work,
 795                    const gmx_edsam      *ed,
 796                    const t_idef         &idef,
 797                    const t_fcdata       *fcd,
 798                    const int             forceFlags
 799                    )
 800 {
 801     forceWork->haveSpecialForces      = haveSpecialForces(inputrec, fr->forceProviders, pull_work, forceFlags, ed);
 802     forceWork->haveCpuBondedWork      = haveCpuBondeds(*fr);
 803     forceWork->haveGpuBondedWork      = ((fr->gpuBonded != nullptr) && fr->gpuBonded->haveInteractions());
 804     forceWork->haveRestraintsWork     = havePositionRestraints(idef, *fcd);
 805     forceWork->haveCpuListedForceWork = haveCpuListedForces(*fr, idef, *fcd);
 806 }
 807
 808
 809 /* \brief Launch end-of-step GPU tasks: buffer clearing and rolling pruning.
 810  *
 811  * TODO: eliminate the \p useGpuNonbonded and \p useGpuNonbonded when these are
 812  * incorporated in PpForceWorkload.
 813  */
 814 static void
 815 launchGpuEndOfStepTasks(nonbonded_verlet_t         *nbv,
 816                         gmx::GpuBonded             *gpuBonded,
 817                         gmx_pme_t                  *pmedata,
 818                         gmx_enerdata_t             *enerd,
 819                         const gmx::PpForceWorkload &forceWorkload,
 820                         bool                        useGpuNonbonded,
 821                         bool                        useGpuPme,
 822                         int64_t                     step,
 823                         int                         flags,
 824                         gmx_wallcycle_t             wcycle)
 825 {
 826     if (useGpuNonbonded)
 827     {
 828         /* Launch pruning before buffer clearing because the API overhead of the
 829          * clear kernel launches can leave the GPU idle while it could be running
 830          * the prune kernel.
 831          */
 832         if (nbv->isDynamicPruningStepGpu(step))
 833         {
 834             wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 835             wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
 836             nbv->dispatchPruneKernelGpu(step);
 837             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
 838             wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 839         }
 840
 841         /* now clear the GPU outputs while we finish the step on the CPU */
 842         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 843         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
 844         Nbnxm::gpu_clear_outputs(nbv->gpu_nbv, flags);
 845         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
 846         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 847     }
 848
 849     if (useGpuPme)
 850     {
 851         pme_gpu_reinit_computation(pmedata, wcycle);
 852     }
 853
 854     if (forceWorkload.haveGpuBondedWork && (flags & GMX_FORCE_ENERGY))
 855     {
 856         // in principle this should be included in the DD balancing region,
 857         // but generally it is infrequent so we'll omit it for the sake of
 858         // simpler code
 859         gpuBonded->waitAccumulateEnergyTerms(enerd);
 860
 861         gpuBonded->clearEnergies();
 862     }
 863 }
 864
 865
 866 void do_force(FILE                                     *fplog,
 867               const t_commrec                          *cr,
 868               const gmx_multisim_t                     *ms,
 869               const t_inputrec                         *inputrec,
 870               gmx::Awh                                 *awh,
 871               gmx_enfrot                               *enforcedRotation,
 872               gmx::ImdSession                          *imdSession,
 873               pull_t                                   *pull_work,
 874               int64_t                                   step,
 875               t_nrnb                                   *nrnb,
 876               gmx_wallcycle_t                           wcycle,
 877               const gmx_localtop_t                     *top,
 878               matrix                                    box,
 879               gmx::ArrayRefWithPadding<gmx::RVec>       x,     //NOLINT(performance-unnecessary-value-param)
 880               history_t                                *hist,
 881               gmx::ArrayRefWithPadding<gmx::RVec>       force, //NOLINT(performance-unnecessary-value-param)
 882               tensor                                    vir_force,
 883               const t_mdatoms                          *mdatoms,
 884               gmx_enerdata_t                           *enerd,
 885               t_fcdata                                 *fcd,
 886               gmx::ArrayRef<real>                       lambda,
 887               t_graph                                  *graph,
 888               t_forcerec                               *fr,
 889               gmx::PpForceWorkload                     *ppForceWorkload,
 890               const gmx_vsite_t                        *vsite,
 891               rvec                                      mu_tot,
 892               double                                    t,
 893               gmx_edsam                                *ed,
 894               int                                       flags,
 895               const DDBalanceRegionHandler             &ddBalanceRegionHandler)
 896 {
 897     int                  i, j;
 898     double               mu[2*DIM];
 899     gmx_bool             bStateChanged, bNS, bFillGrid, bCalcCGCM;
 900     gmx_bool             bDoForces, bUseGPU, bUseOrEmulGPU;
 901     nonbonded_verlet_t  *nbv = fr->nbv.get();
 902     interaction_const_t *ic  = fr->ic;
 903
 904     /* modify force flag if not doing nonbonded */
 905     if (!fr->bNonbonded)
 906     {
 907         flags &= ~GMX_FORCE_NONBONDED;
 908     }
 909     bStateChanged = ((flags & GMX_FORCE_STATECHANGED) != 0);
 910     bNS           = ((flags & GMX_FORCE_NS) != 0);
 911     bFillGrid     = (bNS && bStateChanged);
 912     bCalcCGCM     = (bFillGrid && !DOMAINDECOMP(cr));
 913     bDoForces     = ((flags & GMX_FORCE_FORCES) != 0);
 914     bUseGPU       = fr->nbv->useGpu();
 915     bUseOrEmulGPU = bUseGPU || fr->nbv->emulateGpu();
 916
 917     const auto pmeRunMode = fr->pmedata ? pme_run_mode(fr->pmedata) : PmeRunMode::CPU;
 918     // TODO slim this conditional down - inputrec and duty checks should mean the same in proper code!
 919     const bool useGpuPme  = EEL_PME(fr->ic->eeltype) && thisRankHasDuty(cr, DUTY_PME) &&
 920         ((pmeRunMode == PmeRunMode::GPU) || (pmeRunMode == PmeRunMode::Mixed));
 921     const int  pmeFlags = GMX_PME_SPREAD | GMX_PME_SOLVE |
 922         ((flags & GMX_FORCE_VIRIAL) ? GMX_PME_CALC_ENER_VIR : 0) |
 923         ((flags & GMX_FORCE_ENERGY) ? GMX_PME_CALC_ENER_VIR : 0) |
 924         ((flags & GMX_FORCE_FORCES) ? GMX_PME_CALC_F : 0);
 925     const BufferOpsUseGpu useGpuFBufOps = (c_enableGpuBufOps && bUseGPU && (GMX_GPU == GMX_GPU_CUDA))
 926         && !(flags & (GMX_FORCE_VIRIAL | GMX_FORCE_ENERGY)) ?
 927         BufferOpsUseGpu::True : BufferOpsUseGpu::False;
 928
 929     const BufferOpsUseGpu useGpuXBufOps = (c_enableGpuBufOps && bUseGPU && (GMX_GPU == GMX_GPU_CUDA)) ?
 930         BufferOpsUseGpu::True : BufferOpsUseGpu::False;;
 931
 932     /* At a search step we need to start the first balancing region
 933      * somewhere early inside the step after communication during domain
 934      * decomposition (and not during the previous step as usual).
 935      */
 936     if (bNS)
 937     {
 938         ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::yes);
 939     }
 940
 941     const int start  = 0;
 942     const int homenr = mdatoms->homenr;
 943
 944     clear_mat(vir_force);
 945
 946     if (bStateChanged)
 947     {
 948         if (inputrecNeedMutot(inputrec))
 949         {
 950             /* Calculate total (local) dipole moment in a temporary common array.
 951              * This makes it possible to sum them over nodes faster.
 952              */
 953             calc_mu(start, homenr,
 954                     x.unpaddedArrayRef(), mdatoms->chargeA, mdatoms->chargeB, mdatoms->nChargePerturbed,
 955                     mu, mu+DIM);
 956         }
 957     }
 958
 959     if (fr->ePBC != epbcNONE)
 960     {
 961         /* Compute shift vectors every step,
 962          * because of pressure coupling or box deformation!
 963          */
 964         if ((flags & GMX_FORCE_DYNAMICBOX) && bStateChanged)
 965         {
 966             calc_shifts(box, fr->shift_vec);
 967         }
 968
 969         if (bCalcCGCM)
 970         {
 971             put_atoms_in_box_omp(fr->ePBC, box, x.unpaddedArrayRef().subArray(0, homenr), gmx_omp_nthreads_get(emntDefault));
 972             inc_nrnb(nrnb, eNR_SHIFTX, homenr);
 973         }
 974         else if (EI_ENERGY_MINIMIZATION(inputrec->eI) && graph)
 975         {
 976             unshift_self(graph, box, as_rvec_array(x.unpaddedArrayRef().data()));
 977         }
 978     }
 979
 980     nbnxn_atomdata_copy_shiftvec((flags & GMX_FORCE_DYNAMICBOX) != 0,
 981                                  fr->shift_vec, nbv->nbat.get());
 982
 983 #if GMX_MPI
 984     if (!thisRankHasDuty(cr, DUTY_PME))
 985     {
 986         /* Send particle coordinates to the pme nodes.
 987          * Since this is only implemented for domain decomposition
 988          * and domain decomposition does not use the graph,
 989          * we do not need to worry about shifting.
 990          */
 991         gmx_pme_send_coordinates(cr, box, as_rvec_array(x.unpaddedArrayRef().data()),
 992                                  lambda[efptCOUL], lambda[efptVDW],
 993                                  (flags & (GMX_FORCE_VIRIAL | GMX_FORCE_ENERGY)) != 0,
 994                                  step, wcycle);
 995     }
 996 #endif /* GMX_MPI */
 997
 998     if (useGpuPme)
 999     {
1000         launchPmeGpuSpread(fr->pmedata, box, as_rvec_array(x.unpaddedArrayRef().data()), flags, pmeFlags, wcycle);
1001     }
1002
1003     /* do gridding for pair search */
1004     if (bNS)
1005     {
1006         if (graph && bStateChanged)
1007         {
1008             /* Calculate intramolecular shift vectors to make molecules whole */
1009             mk_mshift(fplog, graph, fr->ePBC, box, as_rvec_array(x.unpaddedArrayRef().data()));
1010         }
1011
1012         // TODO
1013         // - vzero is constant, do we need to pass it?
1014         // - box_diag should be passed directly to nbnxn_put_on_grid
1015         //
1016         rvec vzero;
1017         clear_rvec(vzero);
1018
1019         rvec box_diag;
1020         box_diag[XX] = box[XX][XX];
1021         box_diag[YY] = box[YY][YY];
1022         box_diag[ZZ] = box[ZZ][ZZ];
1023
1024         wallcycle_start(wcycle, ewcNS);
1025         if (!DOMAINDECOMP(cr))
1026         {
1027             wallcycle_sub_start(wcycle, ewcsNBS_GRID_LOCAL);
1028             nbnxn_put_on_grid(nbv, box,
1029                               0, vzero, box_diag,
1030                               nullptr, 0, mdatoms->homenr, -1,
1031                               fr->cginfo, x.unpaddedArrayRef(),
1032                               0, nullptr);
1033             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_LOCAL);
1034         }
1035         else
1036         {
1037             wallcycle_sub_start(wcycle, ewcsNBS_GRID_NONLOCAL);
1038             nbnxn_put_on_grid_nonlocal(nbv, domdec_zones(cr->dd),
1039                                        fr->cginfo, x.unpaddedArrayRef());
1040             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_NONLOCAL);
1041         }
1042
1043         nbv->setAtomProperties(*mdatoms, fr->cginfo);
1044
1045         wallcycle_stop(wcycle, ewcNS);
1046
1047         /* initialize the GPU nbnxm atom data and bonded data structures */
1048         if (bUseGPU)
1049         {
1050             wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1051
1052             wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1053             Nbnxm::gpu_init_atomdata(nbv->gpu_nbv, nbv->nbat.get());
1054             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1055
1056             if (fr->gpuBonded)
1057             {
1058                 /* Now we put all atoms on the grid, we can assign bonded
1059                  * interactions to the GPU, where the grid order is
1060                  * needed. Also the xq, f and fshift device buffers have
1061                  * been reallocated if needed, so the bonded code can
1062                  * learn about them. */
1063                 // TODO the xq, f, and fshift buffers are now shared
1064                 // resources, so they should be maintained by a
1065                 // higher-level object than the nb module.
1066                 fr->gpuBonded->updateInteractionListsAndDeviceBuffers(nbv->getGridIndices(),
1067                                                                       top->idef,
1068                                                                       Nbnxm::gpu_get_xq(nbv->gpu_nbv),
1069                                                                       Nbnxm::gpu_get_f(nbv->gpu_nbv),
1070                                                                       Nbnxm::gpu_get_fshift(nbv->gpu_nbv));
1071             }
1072             wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1073         }
1074
1075         // Need to run after the GPU-offload bonded interaction lists
1076         // are set up to be able to determine whether there is bonded work.
1077         setupForceWorkload(ppForceWorkload,
1078                            inputrec,
1079                            fr,
1080                            pull_work,
1081                            ed,
1082                            top->idef,
1083                            fcd,
1084                            flags);
1085     }
1086
1087     /* do local pair search */
1088     if (bNS)
1089     {
1090         // TODO: fuse this branch with the above bNS block
1091         wallcycle_start_nocount(wcycle, ewcNS);
1092         wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_LOCAL);
1093         /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
1094         nbv->constructPairlist(Nbnxm::InteractionLocality::Local,
1095                                &top->excls, step, nrnb);
1096
1097         nbv->setupGpuShortRangeWork(fr->gpuBonded, Nbnxm::InteractionLocality::Local);
1098
1099         wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_LOCAL);
1100         wallcycle_stop(wcycle, ewcNS);
1101
1102         if (useGpuXBufOps == BufferOpsUseGpu::True)
1103         {
1104             nbv->atomdata_init_copy_x_to_nbat_x_gpu();
1105         }
1106         // For force buffer ops, we use the below conditon rather than
1107         // useGpuFBufOps to ensure that init is performed even if this
1108         // NS step is also a virial step (on which f buf ops are deactivated).
1109         if (c_enableGpuBufOps && bUseGPU && (GMX_GPU == GMX_GPU_CUDA))
1110         {
1111             nbv->atomdata_init_add_nbat_f_to_f_gpu(wcycle);
1112         }
1113     }
1114     else
1115     {
1116         nbv->setCoordinates(Nbnxm::AtomLocality::Local, false,
1117                             x.unpaddedArrayRef(), useGpuXBufOps, pme_gpu_get_device_x(fr->pmedata), wcycle);
1118     }
1119
1120     if (bUseGPU)
1121     {
1122         ddBalanceRegionHandler.openBeforeForceComputationGpu();
1123
1124         wallcycle_start(wcycle, ewcLAUNCH_GPU);
1125
1126         wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1127         Nbnxm::gpu_upload_shiftvec(nbv->gpu_nbv, nbv->nbat.get());
1128         if (bNS || (useGpuXBufOps == BufferOpsUseGpu::False))
1129         {
1130             Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(),
1131                                       Nbnxm::AtomLocality::Local);
1132         }
1133         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1134         // with X buffer ops offloaded to the GPU on all but the search steps
1135
1136         // bonded work not split into separate local and non-local, so with DD
1137         // we can only launch the kernel after non-local coordinates have been received.
1138         if (ppForceWorkload->haveGpuBondedWork && !havePPDomainDecomposition(cr))
1139         {
1140             wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_BONDED);
1141             fr->gpuBonded->launchKernel(fr, flags, box);
1142             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1143         }
1144
1145         /* launch local nonbonded work on GPU */
1146         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1147         do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::Local, enbvClearFNo,
1148                      step, nrnb, wcycle);
1149         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1150         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1151     }
1152
1153     if (useGpuPme)
1154     {
1155         // In PME GPU and mixed mode we launch FFT / gather after the
1156         // X copy/transform to allow overlap as well as after the GPU NB
1157         // launch to avoid FFT launch overhead hijacking the CPU and delaying
1158         // the nonbonded kernel.
1159         launchPmeGpuFftAndGather(fr->pmedata, wcycle);
1160     }
1161
1162     /* Communicate coordinates and sum dipole if necessary +
1163        do non-local pair search */
1164     if (havePPDomainDecomposition(cr))
1165     {
1166         if (bNS)
1167         {
1168             // TODO: fuse this branch with the above large bNS block
1169             wallcycle_start_nocount(wcycle, ewcNS);
1170             wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1171             /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
1172             nbv->constructPairlist(Nbnxm::InteractionLocality::NonLocal,
1173                                    &top->excls, step, nrnb);
1174
1175             nbv->setupGpuShortRangeWork(fr->gpuBonded, Nbnxm::InteractionLocality::NonLocal);
1176             wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1177             wallcycle_stop(wcycle, ewcNS);
1178         }
1179         else
1180         {
1181             dd_move_x(cr->dd, box, x.unpaddedArrayRef(), wcycle);
1182
1183             nbv->setCoordinates(Nbnxm::AtomLocality::NonLocal, false,
1184                                 x.unpaddedArrayRef(), useGpuXBufOps, pme_gpu_get_device_x(fr->pmedata), wcycle);
1185
1186         }
1187
1188         if (bUseGPU)
1189         {
1190             wallcycle_start(wcycle, ewcLAUNCH_GPU);
1191
1192             if (bNS || (useGpuXBufOps == BufferOpsUseGpu::False))
1193             {
1194                 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1195                 Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(),
1196                                           Nbnxm::AtomLocality::NonLocal);
1197                 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1198             }
1199
1200             if (ppForceWorkload->haveGpuBondedWork)
1201             {
1202                 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_BONDED);
1203                 fr->gpuBonded->launchKernel(fr, flags, box);
1204                 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1205             }
1206
1207             /* launch non-local nonbonded tasks on GPU */
1208             wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1209             do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::NonLocal, enbvClearFNo,
1210                          step, nrnb, wcycle);
1211             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1212
1213             wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1214         }
1215     }
1216
1217     if (bUseGPU)
1218     {
1219         /* launch D2H copy-back F */
1220         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1221         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1222
1223         bool copyBackNbForce  = (useGpuFBufOps == BufferOpsUseGpu::False);
1224
1225         if (havePPDomainDecomposition(cr))
1226         {
1227             Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat.get(),
1228                                       flags, Nbnxm::AtomLocality::NonLocal, copyBackNbForce);
1229         }
1230         Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat.get(),
1231                                   flags, Nbnxm::AtomLocality::Local, copyBackNbForce);
1232
1233         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1234
1235         if (ppForceWorkload->haveGpuBondedWork && (flags & GMX_FORCE_ENERGY))
1236         {
1237             fr->gpuBonded->launchEnergyTransfer();
1238         }
1239         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1240     }
1241
1242     if (bStateChanged && inputrecNeedMutot(inputrec))
1243     {
1244         if (PAR(cr))
1245         {
1246             gmx_sumd(2*DIM, mu, cr);
1247
1248             ddBalanceRegionHandler.reopenRegionCpu();
1249         }
1250
1251         for (i = 0; i < 2; i++)
1252         {
1253             for (j = 0; j < DIM; j++)
1254             {
1255                 fr->mu_tot[i][j] = mu[i*DIM + j];
1256             }
1257         }
1258     }
1259     if (fr->efep == efepNO)
1260     {
1261         copy_rvec(fr->mu_tot[0], mu_tot);
1262     }
1263     else
1264     {
1265         for (j = 0; j < DIM; j++)
1266         {
1267             mu_tot[j] =
1268                 (1.0 - lambda[efptCOUL])*fr->mu_tot[0][j] +
1269                 lambda[efptCOUL]*fr->mu_tot[1][j];
1270         }
1271     }
1272
1273     /* Reset energies */
1274     reset_enerdata(enerd);
1275     clear_rvecs(SHIFTS, fr->fshift);
1276
1277     if (DOMAINDECOMP(cr) && !thisRankHasDuty(cr, DUTY_PME))
1278     {
1279         wallcycle_start(wcycle, ewcPPDURINGPME);
1280         dd_force_flop_start(cr->dd, nrnb);
1281     }
1282
1283     if (inputrec->bRot)
1284     {
1285         wallcycle_start(wcycle, ewcROT);
1286         do_rotation(cr, enforcedRotation, box, as_rvec_array(x.unpaddedArrayRef().data()), t, step, bNS);
1287         wallcycle_stop(wcycle, ewcROT);
1288     }
1289
1290     /* Start the force cycle counter.
1291      * Note that a different counter is used for dynamic load balancing.
1292      */
1293     wallcycle_start(wcycle, ewcFORCE);
1294
1295     // set up and clear force outputs
1296     struct ForceOutputs forceOut = setupForceOutputs(fr, pull_work, *inputrec, force, bDoForces,
1297                                                      ((flags & GMX_FORCE_VIRIAL) != 0), wcycle);
1298
1299     /* We calculate the non-bonded forces, when done on the CPU, here.
1300      * We do this before calling do_force_lowlevel, because in that
1301      * function, the listed forces are calculated before PME, which
1302      * does communication.  With this order, non-bonded and listed
1303      * force calculation imbalance can be balanced out by the domain
1304      * decomposition load balancing.
1305      */
1306
1307     if (!bUseOrEmulGPU)
1308     {
1309         do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::Local, enbvClearFYes,
1310                      step, nrnb, wcycle);
1311     }
1312
1313     if (fr->efep != efepNO)
1314     {
1315         /* Calculate the local and non-local free energy interactions here.
1316          * Happens here on the CPU both with and without GPU.
1317          */
1318         nbv->dispatchFreeEnergyKernel(Nbnxm::InteractionLocality::Local,
1319                                       fr, as_rvec_array(x.unpaddedArrayRef().data()), forceOut.f, *mdatoms,
1320                                       inputrec->fepvals, lambda.data(),
1321                                       enerd, flags, nrnb, wcycle);
1322
1323         if (havePPDomainDecomposition(cr))
1324         {
1325             nbv->dispatchFreeEnergyKernel(Nbnxm::InteractionLocality::NonLocal,
1326                                           fr, as_rvec_array(x.unpaddedArrayRef().data()), forceOut.f, *mdatoms,
1327                                           inputrec->fepvals, lambda.data(),
1328                                           enerd, flags, nrnb, wcycle);
1329         }
1330     }
1331
1332     if (!bUseOrEmulGPU)
1333     {
1334         if (havePPDomainDecomposition(cr))
1335         {
1336             do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::NonLocal, enbvClearFNo,
1337                          step, nrnb, wcycle);
1338         }
1339
1340         /* Add all the non-bonded force to the normal force array.
1341          * This can be split into a local and a non-local part when overlapping
1342          * communication with calculation with domain decomposition.
1343          */
1344         wallcycle_stop(wcycle, ewcFORCE);
1345         nbv->atomdata_add_nbat_f_to_f(Nbnxm::AtomLocality::All, forceOut.f,
1346                                       BufferOpsUseGpu::False,
1347                                       GpuBufferOpsAccumulateForce::Null,
1348                                       wcycle);
1349
1350         wallcycle_start_nocount(wcycle, ewcFORCE);
1351
1352         /* If there are multiple fshift output buffers we need to reduce them */
1353         if (flags & GMX_FORCE_VIRIAL)
1354         {
1355             /* This is not in a subcounter because it takes a
1356                negligible and constant-sized amount of time */
1357             nbnxn_atomdata_add_nbat_fshift_to_fshift(nbv->nbat.get(),
1358                                                      fr->fshift);
1359         }
1360     }
1361
1362     /* update QMMMrec, if necessary */
1363     if (fr->bQMMM)
1364     {
1365         update_QMMMrec(cr, fr, as_rvec_array(x.unpaddedArrayRef().data()), mdatoms, box);
1366     }
1367
1368     /* Compute the bonded and non-bonded energies and optionally forces */
1369     do_force_lowlevel(fr, inputrec, &(top->idef),
1370                       cr, ms, nrnb, wcycle, mdatoms,
1371                       x, hist, forceOut.f, &forceOut.forceWithVirial, enerd, fcd,
1372                       box, lambda.data(), graph, fr->mu_tot,
1373                       flags,
1374                       ddBalanceRegionHandler);
1375
1376     wallcycle_stop(wcycle, ewcFORCE);
1377
1378     computeSpecialForces(fplog, cr, inputrec, awh, enforcedRotation,
1379                          imdSession, pull_work, step, t, wcycle,
1380                          fr->forceProviders, box, x.unpaddedArrayRef(), mdatoms, lambda.data(),
1381                          flags, &forceOut.forceWithVirial, enerd,
1382                          ed, bNS);
1383
1384     // flag to specify if CPU force output is preset in force
1385     // buffer. For now, this is true even when useGpuPme == true
1386     // (because on-GPU PME-nonbonded reduction will be added in
1387     // follow-up)
1388     // TODO adapt the below when on-GPU PME-nonbonded reduction is available.
1389     bool                   useCpuPmeReduction = true;
1390     bool                   haveCpuForces      = (ppForceWorkload->haveSpecialForces || ppForceWorkload->haveCpuListedForceWork || useCpuPmeReduction);
1391     // flag to specify if forces should be accumulated in force buffer
1392     // ops. For now, this is solely determined by above haveCpuForces
1393     // flag, but in future developments it will also depend on
1394     // e.g. whether the GPU force halo exchange is active.
1395     GpuBufferOpsAccumulateForce accumulateForce = (useGpuFBufOps == BufferOpsUseGpu::True) &&
1396         haveCpuForces ? GpuBufferOpsAccumulateForce::True :
1397         GpuBufferOpsAccumulateForce::False;
1398
1399     // Will store the amount of cycles spent waiting for the GPU that
1400     // will be later used in the DLB accounting.
1401     float cycles_wait_gpu = 0;
1402     if (bUseOrEmulGPU)
1403     {
1404         /* wait for non-local forces (or calculate in emulation mode) */
1405         if (havePPDomainDecomposition(cr))
1406         {
1407             if (bUseGPU)
1408             {
1409                 wallcycle_start(wcycle, ewcWAIT_GPU_NB_NL);
1410                 Nbnxm::gpu_wait_finish_task(nbv->gpu_nbv,
1411                                             flags, Nbnxm::AtomLocality::NonLocal,
1412                                             enerd->grpp.ener[egLJSR].data(),
1413                                             enerd->grpp.ener[egCOULSR].data(),
1414                                             fr->fshift);
1415                 cycles_wait_gpu += wallcycle_stop(wcycle, ewcWAIT_GPU_NB_NL);
1416             }
1417             else
1418             {
1419                 wallcycle_start_nocount(wcycle, ewcFORCE);
1420                 do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::NonLocal, enbvClearFYes,
1421                              step, nrnb, wcycle);
1422                 wallcycle_stop(wcycle, ewcFORCE);
1423             }
1424
1425             if (useGpuFBufOps == BufferOpsUseGpu::True && haveCpuForces)
1426             {
1427                 nbv->launch_copy_f_to_gpu(forceOut.f, Nbnxm::AtomLocality::NonLocal);
1428             }
1429             nbv->atomdata_add_nbat_f_to_f(Nbnxm::AtomLocality::NonLocal,
1430                                           forceOut.f, useGpuFBufOps, accumulateForce, wcycle);
1431             if (useGpuFBufOps == BufferOpsUseGpu::True)
1432             {
1433                 nbv->launch_copy_f_from_gpu(forceOut.f, Nbnxm::AtomLocality::NonLocal);
1434             }
1435
1436             if (fr->nbv->emulateGpu() && (flags & GMX_FORCE_VIRIAL))
1437             {
1438                 nbnxn_atomdata_add_nbat_fshift_to_fshift(nbv->nbat.get(),
1439                                                          fr->fshift);
1440             }
1441         }
1442     }
1443
1444     if (havePPDomainDecomposition(cr))
1445     {
1446         /* We are done with the CPU compute.
1447          * We will now communicate the non-local forces.
1448          * If we use a GPU this will overlap with GPU work, so in that case
1449          * we do not close the DD force balancing region here.
1450          */
1451         ddBalanceRegionHandler.closeAfterForceComputationCpu();
1452
1453         if (bDoForces)
1454         {
1455             if (useGpuFBufOps == BufferOpsUseGpu::True)
1456             {
1457                 nbv->wait_for_gpu_force_reduction(Nbnxm::AtomLocality::NonLocal);
1458             }
1459             dd_move_f(cr->dd, force.unpaddedArrayRef(), fr->fshift, wcycle);
1460         }
1461     }
1462
1463     // With both nonbonded and PME offloaded a GPU on the same rank, we use
1464     // an alternating wait/reduction scheme.
1465     bool alternateGpuWait = (!c_disableAlternatingWait && useGpuPme && bUseGPU && !DOMAINDECOMP(cr) &&
1466                              (useGpuFBufOps == BufferOpsUseGpu::False));
1467     if (alternateGpuWait)
1468     {
1469         alternatePmeNbGpuWaitReduce(fr->nbv.get(), fr->pmedata, &force, &forceOut.forceWithVirial, fr->fshift, enerd,
1470                                     flags, pmeFlags, wcycle);
1471     }
1472
1473     if (!alternateGpuWait && useGpuPme)
1474     {
1475         pme_gpu_wait_and_reduce(fr->pmedata, pmeFlags, wcycle, &forceOut.forceWithVirial, enerd);
1476     }
1477
1478     /* Wait for local GPU NB outputs on the non-alternating wait path */
1479     if (!alternateGpuWait && bUseGPU)
1480     {
1481         /* Measured overhead on CUDA and OpenCL with(out) GPU sharing
1482          * is between 0.5 and 1.5 Mcycles. So 2 MCycles is an overestimate,
1483          * but even with a step of 0.1 ms the difference is less than 1%
1484          * of the step time.
1485          */
1486         const float gpuWaitApiOverheadMargin = 2e6f; /* cycles */
1487
1488         wallcycle_start(wcycle, ewcWAIT_GPU_NB_L);
1489         Nbnxm::gpu_wait_finish_task(nbv->gpu_nbv,
1490                                     flags, Nbnxm::AtomLocality::Local,
1491                                     enerd->grpp.ener[egLJSR].data(),
1492                                     enerd->grpp.ener[egCOULSR].data(),
1493                                     fr->fshift);
1494         float cycles_tmp = wallcycle_stop(wcycle, ewcWAIT_GPU_NB_L);
1495
1496         if (ddBalanceRegionHandler.useBalancingRegion())
1497         {
1498             DdBalanceRegionWaitedForGpu waitedForGpu = DdBalanceRegionWaitedForGpu::yes;
1499             if (bDoForces && cycles_tmp <= gpuWaitApiOverheadMargin)
1500             {
1501                 /* We measured few cycles, it could be that the kernel
1502                  * and transfer finished earlier and there was no actual
1503                  * wait time, only API call overhead.
1504                  * Then the actual time could be anywhere between 0 and
1505                  * cycles_wait_est. We will use half of cycles_wait_est.
1506                  */
1507                 waitedForGpu = DdBalanceRegionWaitedForGpu::no;
1508             }
1509             ddBalanceRegionHandler.closeAfterForceComputationGpu(cycles_wait_gpu, waitedForGpu);
1510         }
1511     }
1512
1513     if (fr->nbv->emulateGpu())
1514     {
1515         // NOTE: emulation kernel is not included in the balancing region,
1516         // but emulation mode does not target performance anyway
1517         wallcycle_start_nocount(wcycle, ewcFORCE);
1518         do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::Local,
1519                      DOMAINDECOMP(cr) ? enbvClearFNo : enbvClearFYes,
1520                      step, nrnb, wcycle);
1521         wallcycle_stop(wcycle, ewcFORCE);
1522     }
1523
1524     /* Do the nonbonded GPU (or emulation) force buffer reduction
1525      * on the non-alternating path. */
1526     if (bUseOrEmulGPU && !alternateGpuWait)
1527     {
1528
1529         // TODO: move these steps as early as possible:
1530         // - CPU f H2D should be as soon as all CPU-side forces are done
1531         // - wait for force reduction does not need to block host (at least not here, it's sufficient to wait
1532         //   before the next CPU task that consumes the forces: vsite spread or update)
1533         //
1534         if (useGpuFBufOps == BufferOpsUseGpu::True && haveCpuForces)
1535         {
1536             nbv->launch_copy_f_to_gpu(forceOut.f, Nbnxm::AtomLocality::Local);
1537         }
1538         nbv->atomdata_add_nbat_f_to_f(Nbnxm::AtomLocality::Local,
1539                                       forceOut.f, useGpuFBufOps, accumulateForce, wcycle);
1540         if (useGpuFBufOps == BufferOpsUseGpu::True)
1541         {
1542             nbv->launch_copy_f_from_gpu(forceOut.f, Nbnxm::AtomLocality::Local);
1543             nbv->wait_for_gpu_force_reduction(Nbnxm::AtomLocality::Local);
1544         }
1545     }
1546
1547     launchGpuEndOfStepTasks(nbv, fr->gpuBonded, fr->pmedata, enerd,
1548                             *ppForceWorkload,
1549                             bUseGPU, useGpuPme,
1550                             step,
1551                             flags,
1552                             wcycle);
1553
1554     if (DOMAINDECOMP(cr))
1555     {
1556         dd_force_flop_stop(cr->dd, nrnb);
1557     }
1558
1559     if (bDoForces)
1560     {
1561         /* If we have NoVirSum forces, but we do not calculate the virial,
1562          * we sum fr->f_novirsum=forceOut.f later.
1563          */
1564         if (vsite && !(fr->haveDirectVirialContributions && !(flags & GMX_FORCE_VIRIAL)))
1565         {
1566             spread_vsite_f(vsite, as_rvec_array(x.unpaddedArrayRef().data()), forceOut.f, fr->fshift, FALSE, nullptr, nrnb,
1567                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr, wcycle);
1568         }
1569
1570         if (flags & GMX_FORCE_VIRIAL)
1571         {
1572             /* Calculation of the virial must be done after vsites! */
1573             calc_virial(0, mdatoms->homenr, as_rvec_array(x.unpaddedArrayRef().data()), forceOut.f,
1574                         vir_force, graph, box, nrnb, fr, inputrec->ePBC);
1575         }
1576     }
1577
1578     if (PAR(cr) && !thisRankHasDuty(cr, DUTY_PME))
1579     {
1580         /* In case of node-splitting, the PP nodes receive the long-range
1581          * forces, virial and energy from the PME nodes here.
1582          */
1583         pme_receive_force_ener(cr, &forceOut.forceWithVirial, enerd, wcycle);
1584     }
1585
1586     if (bDoForces)
1587     {
1588         post_process_forces(cr, step, nrnb, wcycle,
1589                             top, box, as_rvec_array(x.unpaddedArrayRef().data()), forceOut.f, &forceOut.forceWithVirial,
1590                             vir_force, mdatoms, graph, fr, vsite,
1591                             flags);
1592     }
1593
1594     if (flags & GMX_FORCE_ENERGY)
1595     {
1596         /* Sum the potential energy terms from group contributions */
1597         sum_epot(&(enerd->grpp), enerd->term);
1598
1599         if (!EI_TPI(inputrec->eI))
1600         {
1601             checkPotentialEnergyValidity(step, *enerd, *inputrec);
1602         }
1603     }
1604
1605     /* In case we don't have constraints and are using GPUs, the next balancing
1606      * region starts here.
1607      * Some "special" work at the end of do_force_cuts?, such as vsite spread,
1608      * virial calculation and COM pulling, is not thus not included in
1609      * the balance timing, which is ok as most tasks do communication.
1610      */
1611     ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::no);
1612 }