src/gromacs/mdlib/sim_util.cpp

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