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 "sim_util.h"
  40
  41 #include "config.h"
  42
  43 #include <cmath>
  44 #include <cstdint>
  45 #include <cstdio>
  46 #include <cstring>
  47
  48 #include <array>
  49
  50 #include "gromacs/awh/awh.h"
  51 #include "gromacs/domdec/dlbtiming.h"
  52 #include "gromacs/domdec/domdec.h"
  53 #include "gromacs/domdec/domdec_struct.h"
  54 #include "gromacs/domdec/partition.h"
  55 #include "gromacs/essentialdynamics/edsam.h"
  56 #include "gromacs/ewald/pme.h"
  57 #include "gromacs/gmxlib/chargegroup.h"
  58 #include "gromacs/gmxlib/network.h"
  59 #include "gromacs/gmxlib/nrnb.h"
  60 #include "gromacs/gmxlib/nonbonded/nb_free_energy.h"
  61 #include "gromacs/gmxlib/nonbonded/nb_kernel.h"
  62 #include "gromacs/gmxlib/nonbonded/nonbonded.h"
  63 #include "gromacs/gpu_utils/gpu_utils.h"
  64 #include "gromacs/imd/imd.h"
  65 #include "gromacs/listed_forces/bonded.h"
  66 #include "gromacs/listed_forces/disre.h"
  67 #include "gromacs/listed_forces/gpubonded.h"
  68 #include "gromacs/listed_forces/manage_threading.h"
  69 #include "gromacs/listed_forces/orires.h"
  70 #include "gromacs/math/arrayrefwithpadding.h"
  71 #include "gromacs/math/functions.h"
  72 #include "gromacs/math/units.h"
  73 #include "gromacs/math/vec.h"
  74 #include "gromacs/math/vecdump.h"
  75 #include "gromacs/mdlib/calcmu.h"
  76 #include "gromacs/mdlib/calcvir.h"
  77 #include "gromacs/mdlib/constr.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/mdrun.h"
  82 #include "gromacs/mdlib/ppforceworkload.h"
  83 #include "gromacs/mdlib/qmmm.h"
  84 #include "gromacs/mdlib/update.h"
  85 #include "gromacs/mdtypes/commrec.h"
  86 #include "gromacs/mdtypes/enerdata.h"
  87 #include "gromacs/mdtypes/forceoutput.h"
  88 #include "gromacs/mdtypes/iforceprovider.h"
  89 #include "gromacs/mdtypes/inputrec.h"
  90 #include "gromacs/mdtypes/md_enums.h"
  91 #include "gromacs/mdtypes/state.h"
  92 #include "gromacs/nbnxm/atomdata.h"
  93 #include "gromacs/nbnxm/gpu_data_mgmt.h"
  94 #include "gromacs/nbnxm/nbnxm.h"
  95 #include "gromacs/pbcutil/ishift.h"
  96 #include "gromacs/pbcutil/mshift.h"
  97 #include "gromacs/pbcutil/pbc.h"
  98 #include "gromacs/pulling/pull.h"
  99 #include "gromacs/pulling/pull_rotation.h"
 100 #include "gromacs/timing/cyclecounter.h"
 101 #include "gromacs/timing/gpu_timing.h"
 102 #include "gromacs/timing/wallcycle.h"
 103 #include "gromacs/timing/wallcyclereporting.h"
 104 #include "gromacs/timing/walltime_accounting.h"
 105 #include "gromacs/topology/topology.h"
 106 #include "gromacs/utility/arrayref.h"
 107 #include "gromacs/utility/basedefinitions.h"
 108 #include "gromacs/utility/cstringutil.h"
 109 #include "gromacs/utility/exceptions.h"
 110 #include "gromacs/utility/fatalerror.h"
 111 #include "gromacs/utility/gmxassert.h"
 112 #include "gromacs/utility/gmxmpi.h"
 113 #include "gromacs/utility/logger.h"
 114 #include "gromacs/utility/smalloc.h"
 115 #include "gromacs/utility/strconvert.h"
 116 #include "gromacs/utility/sysinfo.h"
 117
 118 // TODO: this environment variable allows us to verify before release
 119 // that on less common architectures the total cost of polling is not larger than
 120 // a blocking wait (so polling does not introduce overhead when the static
 121 // PME-first ordering would suffice).
 122 static const bool c_disableAlternatingWait = (getenv("GMX_DISABLE_ALTERNATING_GPU_WAIT") != nullptr);
 123
 124
 125 static void sum_forces(rvec f[], gmx::ArrayRef<const gmx::RVec> forceToAdd)
 126 {
 127     const int      end = forceToAdd.size();
 128
 129     int gmx_unused nt = gmx_omp_nthreads_get(emntDefault);
 130 #pragma omp parallel for num_threads(nt) schedule(static)
 131     for (int i = 0; i < end; i++)
 132     {
 133         rvec_inc(f[i], forceToAdd[i]);
 134     }
 135 }
 136
 137 static void calc_virial(int start, int homenr, const rvec x[], const rvec f[],
 138                         tensor vir_part, const t_graph *graph, const matrix box,
 139                         t_nrnb *nrnb, const t_forcerec *fr, int ePBC)
 140 {
 141     /* The short-range virial from surrounding boxes */
 142     calc_vir(SHIFTS, fr->shift_vec, fr->fshift, vir_part, ePBC == epbcSCREW, box);
 143     inc_nrnb(nrnb, eNR_VIRIAL, SHIFTS);
 144
 145     /* Calculate partial virial, for local atoms only, based on short range.
 146      * Total virial is computed in global_stat, called from do_md
 147      */
 148     f_calc_vir(start, start+homenr, x, f, vir_part, graph, box);
 149     inc_nrnb(nrnb, eNR_VIRIAL, homenr);
 150
 151     if (debug)
 152     {
 153         pr_rvecs(debug, 0, "vir_part", vir_part, DIM);
 154     }
 155 }
 156
 157 static void pull_potential_wrapper(const t_commrec *cr,
 158                                    const t_inputrec *ir,
 159                                    const matrix box, gmx::ArrayRef<const gmx::RVec> x,
 160                                    gmx::ForceWithVirial *force,
 161                                    const t_mdatoms *mdatoms,
 162                                    gmx_enerdata_t *enerd,
 163                                    const real *lambda,
 164                                    double t,
 165                                    gmx_wallcycle_t wcycle)
 166 {
 167     t_pbc  pbc;
 168     real   dvdl;
 169
 170     /* Calculate the center of mass forces, this requires communication,
 171      * which is why pull_potential is called close to other communication.
 172      */
 173     wallcycle_start(wcycle, ewcPULLPOT);
 174     set_pbc(&pbc, ir->ePBC, box);
 175     dvdl                     = 0;
 176     enerd->term[F_COM_PULL] +=
 177         pull_potential(ir->pull_work, mdatoms, &pbc,
 178                        cr, t, lambda[efptRESTRAINT], as_rvec_array(x.data()), force, &dvdl);
 179     enerd->dvdl_lin[efptRESTRAINT] += dvdl;
 180     wallcycle_stop(wcycle, ewcPULLPOT);
 181 }
 182
 183 static void pme_receive_force_ener(const t_commrec      *cr,
 184                                    gmx::ForceWithVirial *forceWithVirial,
 185                                    gmx_enerdata_t       *enerd,
 186                                    gmx_wallcycle_t       wcycle)
 187 {
 188     real   e_q, e_lj, dvdl_q, dvdl_lj;
 189     float  cycles_ppdpme, cycles_seppme;
 190
 191     cycles_ppdpme = wallcycle_stop(wcycle, ewcPPDURINGPME);
 192     dd_cycles_add(cr->dd, cycles_ppdpme, ddCyclPPduringPME);
 193
 194     /* In case of node-splitting, the PP nodes receive the long-range
 195      * forces, virial and energy from the PME nodes here.
 196      */
 197     wallcycle_start(wcycle, ewcPP_PMEWAITRECVF);
 198     dvdl_q  = 0;
 199     dvdl_lj = 0;
 200     gmx_pme_receive_f(cr, forceWithVirial, &e_q, &e_lj, &dvdl_q, &dvdl_lj,
 201                       &cycles_seppme);
 202     enerd->term[F_COUL_RECIP] += e_q;
 203     enerd->term[F_LJ_RECIP]   += e_lj;
 204     enerd->dvdl_lin[efptCOUL] += dvdl_q;
 205     enerd->dvdl_lin[efptVDW]  += dvdl_lj;
 206
 207     if (wcycle)
 208     {
 209         dd_cycles_add(cr->dd, cycles_seppme, ddCyclPME);
 210     }
 211     wallcycle_stop(wcycle, ewcPP_PMEWAITRECVF);
 212 }
 213
 214 static void print_large_forces(FILE            *fp,
 215                                const t_mdatoms *md,
 216                                const t_commrec *cr,
 217                                int64_t          step,
 218                                real             forceTolerance,
 219                                const rvec      *x,
 220                                const rvec      *f)
 221 {
 222     real           force2Tolerance = gmx::square(forceTolerance);
 223     gmx::index     numNonFinite    = 0;
 224     for (int i = 0; i < md->homenr; i++)
 225     {
 226         real force2    = norm2(f[i]);
 227         bool nonFinite = !std::isfinite(force2);
 228         if (force2 >= force2Tolerance || nonFinite)
 229         {
 230             fprintf(fp, "step %" PRId64 " atom %6d  x %8.3f %8.3f %8.3f  force %12.5e\n",
 231                     step,
 232                     ddglatnr(cr->dd, i), x[i][XX], x[i][YY], x[i][ZZ], std::sqrt(force2));
 233         }
 234         if (nonFinite)
 235         {
 236             numNonFinite++;
 237         }
 238     }
 239     if (numNonFinite > 0)
 240     {
 241         /* Note that with MPI this fatal call on one rank might interrupt
 242          * the printing on other ranks. But we can only avoid that with
 243          * an expensive MPI barrier that we would need at each step.
 244          */
 245         gmx_fatal(FARGS, "At step %" PRId64 " detected non-finite forces on %td atoms", step, numNonFinite);
 246     }
 247 }
 248
 249 static void post_process_forces(const t_commrec           *cr,
 250                                 int64_t                    step,
 251                                 t_nrnb                    *nrnb,
 252                                 gmx_wallcycle_t            wcycle,
 253                                 const gmx_localtop_t      *top,
 254                                 const matrix               box,
 255                                 const rvec                 x[],
 256                                 rvec                       f[],
 257                                 gmx::ForceWithVirial      *forceWithVirial,
 258                                 tensor                     vir_force,
 259                                 const t_mdatoms           *mdatoms,
 260                                 const t_graph             *graph,
 261                                 const t_forcerec          *fr,
 262                                 const gmx_vsite_t         *vsite,
 263                                 int                        flags)
 264 {
 265     if (fr->haveDirectVirialContributions)
 266     {
 267         rvec *fDirectVir = as_rvec_array(forceWithVirial->force_.data());
 268
 269         if (vsite)
 270         {
 271             /* Spread the mesh force on virtual sites to the other particles...
 272              * This is parallellized. MPI communication is performed
 273              * if the constructing atoms aren't local.
 274              */
 275             matrix virial = { { 0 } };
 276             spread_vsite_f(vsite, x, fDirectVir, nullptr,
 277                            (flags & GMX_FORCE_VIRIAL) != 0, virial,
 278                            nrnb,
 279                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr, wcycle);
 280             forceWithVirial->addVirialContribution(virial);
 281         }
 282
 283         if (flags & GMX_FORCE_VIRIAL)
 284         {
 285             /* Now add the forces, this is local */
 286             sum_forces(f, forceWithVirial->force_);
 287
 288             /* Add the direct virial contributions */
 289             GMX_ASSERT(forceWithVirial->computeVirial_, "forceWithVirial should request virial computation when we request the virial");
 290             m_add(vir_force, forceWithVirial->getVirial(), vir_force);
 291
 292             if (debug)
 293             {
 294                 pr_rvecs(debug, 0, "vir_force", vir_force, DIM);
 295             }
 296         }
 297     }
 298
 299     if (fr->print_force >= 0)
 300     {
 301         print_large_forces(stderr, mdatoms, cr, step, fr->print_force, x, f);
 302     }
 303 }
 304
 305 static void do_nb_verlet(t_forcerec                       *fr,
 306                          const interaction_const_t        *ic,
 307                          gmx_enerdata_t                   *enerd,
 308                          const int                         flags,
 309                          const Nbnxm::InteractionLocality  ilocality,
 310                          const int                         clearF,
 311                          const int64_t                     step,
 312                          t_nrnb                           *nrnb,
 313                          gmx_wallcycle_t                   wcycle)
 314 {
 315     if (!(flags & GMX_FORCE_NONBONDED))
 316     {
 317         /* skip non-bonded calculation */
 318         return;
 319     }
 320
 321     nonbonded_verlet_t *nbv  = fr->nbv;
 322
 323     /* GPU kernel launch overhead is already timed separately */
 324     if (fr->cutoff_scheme != ecutsVERLET)
 325     {
 326         gmx_incons("Invalid cut-off scheme passed!");
 327     }
 328
 329     if (!nbv->useGpu())
 330     {
 331         /* When dynamic pair-list  pruning is requested, we need to prune
 332          * at nstlistPrune steps.
 333          */
 334         if (nbv->pairlistSets().isDynamicPruningStepCpu(step))
 335         {
 336             /* Prune the pair-list beyond fr->ic->rlistPrune using
 337              * the current coordinates of the atoms.
 338              */
 339             wallcycle_sub_start(wcycle, ewcsNONBONDED_PRUNING);
 340             nbv->dispatchPruneKernelCpu(ilocality, fr->shift_vec);
 341             wallcycle_sub_stop(wcycle, ewcsNONBONDED_PRUNING);
 342         }
 343
 344         wallcycle_sub_start(wcycle, ewcsNONBONDED);
 345     }
 346
 347     nbv->dispatchNonbondedKernel(ilocality, *ic, flags, clearF, fr, enerd, nrnb);
 348
 349     if (!nbv->useGpu())
 350     {
 351         wallcycle_sub_stop(wcycle, ewcsNONBONDED);
 352     }
 353 }
 354
 355 gmx_bool use_GPU(const nonbonded_verlet_t *nbv)
 356 {
 357     return nbv != nullptr && nbv->useGpu();
 358 }
 359
 360 static inline void clear_rvecs_omp(int n, rvec v[])
 361 {
 362     int nth = gmx_omp_nthreads_get_simple_rvec_task(emntDefault, n);
 363
 364     /* Note that we would like to avoid this conditional by putting it
 365      * into the omp pragma instead, but then we still take the full
 366      * omp parallel for overhead (at least with gcc5).
 367      */
 368     if (nth == 1)
 369     {
 370         for (int i = 0; i < n; i++)
 371         {
 372             clear_rvec(v[i]);
 373         }
 374     }
 375     else
 376     {
 377 #pragma omp parallel for num_threads(nth) schedule(static)
 378         for (int i = 0; i < n; i++)
 379         {
 380             clear_rvec(v[i]);
 381         }
 382     }
 383 }
 384
 385 /*! \brief Return an estimate of the average kinetic energy or 0 when unreliable
 386  *
 387  * \param groupOptions  Group options, containing T-coupling options
 388  */
 389 static real averageKineticEnergyEstimate(const t_grpopts &groupOptions)
 390 {
 391     real nrdfCoupled   = 0;
 392     real nrdfUncoupled = 0;
 393     real kineticEnergy = 0;
 394     for (int g = 0; g < groupOptions.ngtc; g++)
 395     {
 396         if (groupOptions.tau_t[g] >= 0)
 397         {
 398             nrdfCoupled   += groupOptions.nrdf[g];
 399             kineticEnergy += groupOptions.nrdf[g]*0.5*groupOptions.ref_t[g]*BOLTZ;
 400         }
 401         else
 402         {
 403             nrdfUncoupled += groupOptions.nrdf[g];
 404         }
 405     }
 406
 407     /* This conditional with > also catches nrdf=0 */
 408     if (nrdfCoupled > nrdfUncoupled)
 409     {
 410         return kineticEnergy*(nrdfCoupled + nrdfUncoupled)/nrdfCoupled;
 411     }
 412     else
 413     {
 414         return 0;
 415     }
 416 }
 417
 418 /*! \brief This routine checks that the potential energy is finite.
 419  *
 420  * Always checks that the potential energy is finite. If step equals
 421  * inputrec.init_step also checks that the magnitude of the potential energy
 422  * is reasonable. Terminates with a fatal error when a check fails.
 423  * Note that passing this check does not guarantee finite forces,
 424  * since those use slightly different arithmetics. But in most cases
 425  * there is just a narrow coordinate range where forces are not finite
 426  * and energies are finite.
 427  *
 428  * \param[in] step      The step number, used for checking and printing
 429  * \param[in] enerd     The energy data; the non-bonded group energies need to be added to enerd.term[F_EPOT] before calling this routine
 430  * \param[in] inputrec  The input record
 431  */
 432 static void checkPotentialEnergyValidity(int64_t               step,
 433                                          const gmx_enerdata_t &enerd,
 434                                          const t_inputrec     &inputrec)
 435 {
 436     /* Threshold valid for comparing absolute potential energy against
 437      * the kinetic energy. Normally one should not consider absolute
 438      * potential energy values, but with a factor of one million
 439      * we should never get false positives.
 440      */
 441     constexpr real c_thresholdFactor = 1e6;
 442
 443     bool           energyIsNotFinite    = !std::isfinite(enerd.term[F_EPOT]);
 444     real           averageKineticEnergy = 0;
 445     /* We only check for large potential energy at the initial step,
 446      * because that is by far the most likely step for this too occur
 447      * and because computing the average kinetic energy is not free.
 448      * Note: nstcalcenergy >> 1 often does not allow to catch large energies
 449      * before they become NaN.
 450      */
 451     if (step == inputrec.init_step && EI_DYNAMICS(inputrec.eI))
 452     {
 453         averageKineticEnergy = averageKineticEnergyEstimate(inputrec.opts);
 454     }
 455
 456     if (energyIsNotFinite || (averageKineticEnergy > 0 &&
 457                               enerd.term[F_EPOT] > c_thresholdFactor*averageKineticEnergy))
 458     {
 459         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.",
 460                   step,
 461                   enerd.term[F_EPOT],
 462                   energyIsNotFinite ? "not finite" : "extremely high",
 463                   enerd.term[F_LJ],
 464                   enerd.term[F_COUL_SR],
 465                   energyIsNotFinite ? "non-finite" : "very high",
 466                   energyIsNotFinite ? " or Nan" : "");
 467     }
 468 }
 469
 470 /*! \brief Compute forces and/or energies for special algorithms
 471  *
 472  * The intention is to collect all calls to algorithms that compute
 473  * forces on local atoms only and that do not contribute to the local
 474  * virial sum (but add their virial contribution separately).
 475  * Eventually these should likely all become ForceProviders.
 476  * Within this function the intention is to have algorithms that do
 477  * global communication at the end, so global barriers within the MD loop
 478  * are as close together as possible.
 479  *
 480  * \param[in]     fplog            The log file
 481  * \param[in]     cr               The communication record
 482  * \param[in]     inputrec         The input record
 483  * \param[in]     awh              The Awh module (nullptr if none in use).
 484  * \param[in]     enforcedRotation Enforced rotation module.
 485  * \param[in]     step             The current MD step
 486  * \param[in]     t                The current time
 487  * \param[in,out] wcycle           Wallcycle accounting struct
 488  * \param[in,out] forceProviders   Pointer to a list of force providers
 489  * \param[in]     box              The unit cell
 490  * \param[in]     x                The coordinates
 491  * \param[in]     mdatoms          Per atom properties
 492  * \param[in]     lambda           Array of free-energy lambda values
 493  * \param[in]     forceFlags       Flags that tell whether we should compute forces/energies/virial
 494  * \param[in,out] forceWithVirial  Force and virial buffers
 495  * \param[in,out] enerd            Energy buffer
 496  * \param[in,out] ed               Essential dynamics pointer
 497  * \param[in]     bNS              Tells if we did neighbor searching this step, used for ED sampling
 498  *
 499  * \todo Remove bNS, which is used incorrectly.
 500  * \todo Convert all other algorithms called here to ForceProviders.
 501  */
 502 static void
 503 computeSpecialForces(FILE                          *fplog,
 504                      const t_commrec               *cr,
 505                      const t_inputrec              *inputrec,
 506                      gmx::Awh                      *awh,
 507                      gmx_enfrot                    *enforcedRotation,
 508                      int64_t                        step,
 509                      double                         t,
 510                      gmx_wallcycle_t                wcycle,
 511                      ForceProviders                *forceProviders,
 512                      matrix                         box,
 513                      gmx::ArrayRef<const gmx::RVec> x,
 514                      const t_mdatoms               *mdatoms,
 515                      real                          *lambda,
 516                      int                            forceFlags,
 517                      gmx::ForceWithVirial          *forceWithVirial,
 518                      gmx_enerdata_t                *enerd,
 519                      gmx_edsam                     *ed,
 520                      gmx_bool                       bNS)
 521 {
 522     const bool computeForces = (forceFlags & GMX_FORCE_FORCES) != 0;
 523
 524     /* NOTE: Currently all ForceProviders only provide forces.
 525      *       When they also provide energies, remove this conditional.
 526      */
 527     if (computeForces)
 528     {
 529         gmx::ForceProviderInput  forceProviderInput(x, *mdatoms, t, box, *cr);
 530         gmx::ForceProviderOutput forceProviderOutput(forceWithVirial, enerd);
 531
 532         /* Collect forces from modules */
 533         forceProviders->calculateForces(forceProviderInput, &forceProviderOutput);
 534     }
 535
 536     if (inputrec->bPull && pull_have_potential(inputrec->pull_work))
 537     {
 538         pull_potential_wrapper(cr, inputrec, box, x,
 539                                forceWithVirial,
 540                                mdatoms, enerd, lambda, t,
 541                                wcycle);
 542
 543         if (awh)
 544         {
 545             enerd->term[F_COM_PULL] +=
 546                 awh->applyBiasForcesAndUpdateBias(inputrec->ePBC, *mdatoms, box,
 547                                                   forceWithVirial,
 548                                                   t, step, wcycle, fplog);
 549         }
 550     }
 551
 552     rvec *f = as_rvec_array(forceWithVirial->force_.data());
 553
 554     /* Add the forces from enforced rotation potentials (if any) */
 555     if (inputrec->bRot)
 556     {
 557         wallcycle_start(wcycle, ewcROTadd);
 558         enerd->term[F_COM_PULL] += add_rot_forces(enforcedRotation, f, cr, step, t);
 559         wallcycle_stop(wcycle, ewcROTadd);
 560     }
 561
 562     if (ed)
 563     {
 564         /* Note that since init_edsam() is called after the initialization
 565          * of forcerec, edsam doesn't request the noVirSum force buffer.
 566          * Thus if no other algorithm (e.g. PME) requires it, the forces
 567          * here will contribute to the virial.
 568          */
 569         do_flood(cr, inputrec, as_rvec_array(x.data()), f, ed, box, step, bNS);
 570     }
 571
 572     /* Add forces from interactive molecular dynamics (IMD), if bIMD == TRUE. */
 573     if (inputrec->bIMD && computeForces)
 574     {
 575         IMD_apply_forces(inputrec->bIMD, inputrec->imd, cr, f, wcycle);
 576     }
 577 }
 578
 579 /*! \brief Launch the prepare_step and spread stages of PME GPU.
 580  *
 581  * \param[in]  pmedata       The PME structure
 582  * \param[in]  box           The box matrix
 583  * \param[in]  x             Coordinate array
 584  * \param[in]  flags         Force flags
 585  * \param[in]  pmeFlags      PME flags
 586  * \param[in]  wcycle        The wallcycle structure
 587  */
 588 static inline void launchPmeGpuSpread(gmx_pme_t      *pmedata,
 589                                       matrix          box,
 590                                       rvec            x[],
 591                                       int             flags,
 592                                       int             pmeFlags,
 593                                       gmx_wallcycle_t wcycle)
 594 {
 595     pme_gpu_prepare_computation(pmedata, (flags & GMX_FORCE_DYNAMICBOX) != 0, box, wcycle, pmeFlags);
 596     pme_gpu_launch_spread(pmedata, x, wcycle);
 597 }
 598
 599 /*! \brief Launch the FFT and gather stages of PME GPU
 600  *
 601  * This function only implements setting the output forces (no accumulation).
 602  *
 603  * \param[in]  pmedata        The PME structure
 604  * \param[in]  wcycle         The wallcycle structure
 605  */
 606 static void launchPmeGpuFftAndGather(gmx_pme_t        *pmedata,
 607                                      gmx_wallcycle_t   wcycle)
 608 {
 609     pme_gpu_launch_complex_transforms(pmedata, wcycle);
 610     pme_gpu_launch_gather(pmedata, wcycle, PmeForceOutputHandling::Set);
 611 }
 612
 613 /*! \brief
 614  *  Polling wait for either of the PME or nonbonded GPU tasks.
 615  *
 616  * Instead of a static order in waiting for GPU tasks, this function
 617  * polls checking which of the two tasks completes first, and does the
 618  * associated force buffer reduction overlapped with the other task.
 619  * By doing that, unlike static scheduling order, it can always overlap
 620  * one of the reductions, regardless of the GPU task completion order.
 621  *
 622  * \param[in]     nbv              Nonbonded verlet structure
 623  * \param[in,out] pmedata          PME module data
 624  * \param[in,out] force            Force array to reduce task outputs into.
 625  * \param[in,out] forceWithVirial  Force and virial buffers
 626  * \param[in,out] fshift           Shift force output vector results are reduced into
 627  * \param[in,out] enerd            Energy data structure results are reduced into
 628  * \param[in]     flags            Force flags
 629  * \param[in]     pmeFlags         PME flags
 630  * \param[in]     haveOtherWork    Tells whether there is other work than non-bonded in the stream(s)
 631  * \param[in]     wcycle           The wallcycle structure
 632  */
 633 static void alternatePmeNbGpuWaitReduce(nonbonded_verlet_t                  *nbv,
 634                                         gmx_pme_t                           *pmedata,
 635                                         gmx::ArrayRefWithPadding<gmx::RVec> *force,
 636                                         gmx::ForceWithVirial                *forceWithVirial,
 637                                         rvec                                 fshift[],
 638                                         gmx_enerdata_t                      *enerd,
 639                                         int                                  flags,
 640                                         int                                  pmeFlags,
 641                                         bool                                 haveOtherWork,
 642                                         gmx_wallcycle_t                      wcycle)
 643 {
 644     bool isPmeGpuDone = false;
 645     bool isNbGpuDone  = false;
 646
 647
 648     gmx::ArrayRef<const gmx::RVec> pmeGpuForces;
 649
 650     while (!isPmeGpuDone || !isNbGpuDone)
 651     {
 652         if (!isPmeGpuDone)
 653         {
 654             GpuTaskCompletion completionType = (isNbGpuDone) ? GpuTaskCompletion::Wait : GpuTaskCompletion::Check;
 655             isPmeGpuDone = pme_gpu_try_finish_task(pmedata, pmeFlags, wcycle, forceWithVirial, enerd, completionType);
 656         }
 657
 658         if (!isNbGpuDone)
 659         {
 660             GpuTaskCompletion completionType = (isPmeGpuDone) ? GpuTaskCompletion::Wait : GpuTaskCompletion::Check;
 661             wallcycle_start_nocount(wcycle, ewcWAIT_GPU_NB_L);
 662             isNbGpuDone = Nbnxm::gpu_try_finish_task(nbv->gpu_nbv,
 663                                                      flags,
 664                                                      Nbnxm::AtomLocality::Local,
 665                                                      haveOtherWork,
 666                                                      enerd->grpp.ener[egLJSR], enerd->grpp.ener[egCOULSR],
 667                                                      fshift, completionType);
 668             wallcycle_stop(wcycle, ewcWAIT_GPU_NB_L);
 669             // To get the call count right, when the task finished we
 670             // issue a start/stop.
 671             // TODO: move the ewcWAIT_GPU_NB_L cycle counting into nbnxn_gpu_try_finish_task()
 672             // and ewcNB_XF_BUF_OPS counting into nbnxn_atomdata_add_nbat_f_to_f().
 673             if (isNbGpuDone)
 674             {
 675                 wallcycle_start(wcycle, ewcWAIT_GPU_NB_L);
 676                 wallcycle_stop(wcycle, ewcWAIT_GPU_NB_L);
 677
 678                 nbv->atomdata_add_nbat_f_to_f(Nbnxm::AtomLocality::Local,
 679                                               as_rvec_array(force->unpaddedArrayRef().data()), wcycle);
 680             }
 681         }
 682     }
 683 }
 684
 685 static void do_force_cutsVERLET(FILE *fplog,
 686                                 const t_commrec *cr,
 687                                 const gmx_multisim_t *ms,
 688                                 const t_inputrec *inputrec,
 689                                 gmx::Awh *awh,
 690                                 gmx_enfrot *enforcedRotation,
 691                                 int64_t step,
 692                                 t_nrnb *nrnb,
 693                                 gmx_wallcycle_t wcycle,
 694                                 const gmx_localtop_t *top,
 695                                 const gmx_groups_t * /* groups */,
 696                                 matrix box, gmx::ArrayRefWithPadding<gmx::RVec> x,
 697                                 history_t *hist,
 698                                 gmx::ArrayRefWithPadding<gmx::RVec> force,
 699                                 tensor vir_force,
 700                                 const t_mdatoms *mdatoms,
 701                                 gmx_enerdata_t *enerd, t_fcdata *fcd,
 702                                 real *lambda,
 703                                 t_graph *graph,
 704                                 t_forcerec *fr,
 705                                 gmx::PpForceWorkload *ppForceWorkload,
 706                                 interaction_const_t *ic,
 707                                 const gmx_vsite_t *vsite,
 708                                 rvec mu_tot,
 709                                 double t,
 710                                 gmx_edsam *ed,
 711                                 const int flags,
 712                                 const DDBalanceRegionHandler &ddBalanceRegionHandler)
 713 {
 714     int                 cg1, i, j;
 715     double              mu[2*DIM];
 716     gmx_bool            bStateChanged, bNS, bFillGrid, bCalcCGCM;
 717     gmx_bool            bDoForces, bUseGPU, bUseOrEmulGPU;
 718     rvec                vzero, box_diag;
 719     float               cycles_pme, cycles_wait_gpu;
 720     nonbonded_verlet_t *nbv = fr->nbv;
 721
 722     bStateChanged = ((flags & GMX_FORCE_STATECHANGED) != 0);
 723     bNS           = ((flags & GMX_FORCE_NS) != 0);
 724     bFillGrid     = (bNS && bStateChanged);
 725     bCalcCGCM     = (bFillGrid && !DOMAINDECOMP(cr));
 726     bDoForces     = ((flags & GMX_FORCE_FORCES) != 0);
 727     bUseGPU       = fr->nbv->useGpu();
 728     bUseOrEmulGPU = bUseGPU || fr->nbv->emulateGpu();
 729
 730     const auto pmeRunMode = fr->pmedata ? pme_run_mode(fr->pmedata) : PmeRunMode::CPU;
 731     // TODO slim this conditional down - inputrec and duty checks should mean the same in proper code!
 732     const bool useGpuPme  = EEL_PME(fr->ic->eeltype) && thisRankHasDuty(cr, DUTY_PME) &&
 733         ((pmeRunMode == PmeRunMode::GPU) || (pmeRunMode == PmeRunMode::Mixed));
 734     const int  pmeFlags = GMX_PME_SPREAD | GMX_PME_SOLVE |
 735         ((flags & GMX_FORCE_VIRIAL) ? GMX_PME_CALC_ENER_VIR : 0) |
 736         ((flags & GMX_FORCE_ENERGY) ? GMX_PME_CALC_ENER_VIR : 0) |
 737         ((flags & GMX_FORCE_FORCES) ? GMX_PME_CALC_F : 0);
 738
 739     /* At a search step we need to start the first balancing region
 740      * somewhere early inside the step after communication during domain
 741      * decomposition (and not during the previous step as usual).
 742      */
 743     if (bNS)
 744     {
 745         ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::yes);
 746     }
 747
 748     cycles_wait_gpu = 0;
 749
 750     const int start  = 0;
 751     const int homenr = mdatoms->homenr;
 752
 753     clear_mat(vir_force);
 754
 755     if (DOMAINDECOMP(cr))
 756     {
 757         cg1 = cr->dd->globalAtomGroupIndices.size();
 758     }
 759     else
 760     {
 761         cg1 = top->cgs.nr;
 762     }
 763     if (fr->n_tpi > 0)
 764     {
 765         cg1--;
 766     }
 767
 768     if (bStateChanged)
 769     {
 770         update_forcerec(fr, box);
 771
 772         if (inputrecNeedMutot(inputrec))
 773         {
 774             /* Calculate total (local) dipole moment in a temporary common array.
 775              * This makes it possible to sum them over nodes faster.
 776              */
 777             calc_mu(start, homenr,
 778                     x.unpaddedArrayRef(), mdatoms->chargeA, mdatoms->chargeB, mdatoms->nChargePerturbed,
 779                     mu, mu+DIM);
 780         }
 781     }
 782
 783     if (fr->ePBC != epbcNONE)
 784     {
 785         /* Compute shift vectors every step,
 786          * because of pressure coupling or box deformation!
 787          */
 788         if ((flags & GMX_FORCE_DYNAMICBOX) && bStateChanged)
 789         {
 790             calc_shifts(box, fr->shift_vec);
 791         }
 792
 793         if (bCalcCGCM)
 794         {
 795             put_atoms_in_box_omp(fr->ePBC, box, x.unpaddedArrayRef().subArray(0, homenr));
 796             inc_nrnb(nrnb, eNR_SHIFTX, homenr);
 797         }
 798         else if (EI_ENERGY_MINIMIZATION(inputrec->eI) && graph)
 799         {
 800             unshift_self(graph, box, as_rvec_array(x.unpaddedArrayRef().data()));
 801         }
 802     }
 803
 804     nbnxn_atomdata_copy_shiftvec((flags & GMX_FORCE_DYNAMICBOX) != 0,
 805                                  fr->shift_vec, nbv->nbat);
 806
 807 #if GMX_MPI
 808     if (!thisRankHasDuty(cr, DUTY_PME))
 809     {
 810         /* Send particle coordinates to the pme nodes.
 811          * Since this is only implemented for domain decomposition
 812          * and domain decomposition does not use the graph,
 813          * we do not need to worry about shifting.
 814          */
 815         gmx_pme_send_coordinates(cr, box, as_rvec_array(x.unpaddedArrayRef().data()),
 816                                  lambda[efptCOUL], lambda[efptVDW],
 817                                  (flags & (GMX_FORCE_VIRIAL | GMX_FORCE_ENERGY)) != 0,
 818                                  step, wcycle);
 819     }
 820 #endif /* GMX_MPI */
 821
 822     if (useGpuPme)
 823     {
 824         launchPmeGpuSpread(fr->pmedata, box, as_rvec_array(x.unpaddedArrayRef().data()), flags, pmeFlags, wcycle);
 825     }
 826
 827     /* do gridding for pair search */
 828     if (bNS)
 829     {
 830         if (graph && bStateChanged)
 831         {
 832             /* Calculate intramolecular shift vectors to make molecules whole */
 833             mk_mshift(fplog, graph, fr->ePBC, box, as_rvec_array(x.unpaddedArrayRef().data()));
 834         }
 835
 836         clear_rvec(vzero);
 837         box_diag[XX] = box[XX][XX];
 838         box_diag[YY] = box[YY][YY];
 839         box_diag[ZZ] = box[ZZ][ZZ];
 840
 841         wallcycle_start(wcycle, ewcNS);
 842         if (!DOMAINDECOMP(cr))
 843         {
 844             wallcycle_sub_start(wcycle, ewcsNBS_GRID_LOCAL);
 845             nbnxn_put_on_grid(nbv, box,
 846                               0, vzero, box_diag,
 847                               nullptr, 0, mdatoms->homenr, -1,
 848                               fr->cginfo, x.unpaddedArrayRef(),
 849                               0, nullptr);
 850             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_LOCAL);
 851         }
 852         else
 853         {
 854             wallcycle_sub_start(wcycle, ewcsNBS_GRID_NONLOCAL);
 855             nbnxn_put_on_grid_nonlocal(nbv, domdec_zones(cr->dd),
 856                                        fr->cginfo, x.unpaddedArrayRef());
 857             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_NONLOCAL);
 858         }
 859
 860         nbnxn_atomdata_set(nbv->nbat, nbv->nbs.get(), mdatoms, fr->cginfo);
 861
 862         wallcycle_stop(wcycle, ewcNS);
 863     }
 864
 865     /* initialize the GPU atom data and copy shift vector */
 866     if (bUseGPU)
 867     {
 868         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 869         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
 870
 871         if (bNS)
 872         {
 873             Nbnxm::gpu_init_atomdata(nbv->gpu_nbv, nbv->nbat);
 874         }
 875
 876         Nbnxm::gpu_upload_shiftvec(nbv->gpu_nbv, nbv->nbat);
 877
 878         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
 879
 880         if (bNS && fr->gpuBonded)
 881         {
 882             /* Now we put all atoms on the grid, we can assign bonded
 883              * interactions to the GPU, where the grid order is
 884              * needed. Also the xq, f and fshift device buffers have
 885              * been reallocated if needed, so the bonded code can
 886              * learn about them. */
 887             // TODO the xq, f, and fshift buffers are now shared
 888             // resources, so they should be maintained by a
 889             // higher-level object than the nb module.
 890             fr->gpuBonded->updateInteractionListsAndDeviceBuffers(nbnxn_get_gridindices(fr->nbv->nbs.get()),
 891                                                                   top->idef,
 892                                                                   Nbnxm::gpu_get_xq(nbv->gpu_nbv),
 893                                                                   Nbnxm::gpu_get_f(nbv->gpu_nbv),
 894                                                                   Nbnxm::gpu_get_fshift(nbv->gpu_nbv));
 895             ppForceWorkload->haveGpuBondedWork = fr->gpuBonded->haveInteractions();
 896         }
 897
 898         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 899     }
 900
 901     /* do local pair search */
 902     if (bNS)
 903     {
 904         wallcycle_start_nocount(wcycle, ewcNS);
 905         wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_LOCAL);
 906         /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
 907         nbv->constructPairlist(Nbnxm::InteractionLocality::Local,
 908                                &top->excls, step, nrnb);
 909         wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_LOCAL);
 910         wallcycle_stop(wcycle, ewcNS);
 911     }
 912     else
 913     {
 914         nbnxn_atomdata_copy_x_to_nbat_x(nbv->nbs.get(), Nbnxm::AtomLocality::Local,
 915                                         FALSE, as_rvec_array(x.unpaddedArrayRef().data()),
 916                                         nbv->nbat, wcycle);
 917     }
 918
 919     if (bUseGPU)
 920     {
 921         ddBalanceRegionHandler.openBeforeForceComputationGpu();
 922
 923         wallcycle_start(wcycle, ewcLAUNCH_GPU);
 924
 925         wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
 926         Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat, Nbnxm::AtomLocality::Local, ppForceWorkload->haveGpuBondedWork);
 927         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
 928
 929         // bonded work not split into separate local and non-local, so with DD
 930         // we can only launch the kernel after non-local coordinates have been received.
 931         if (ppForceWorkload->haveGpuBondedWork && !havePPDomainDecomposition(cr))
 932         {
 933             wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_BONDED);
 934             fr->gpuBonded->launchKernels(fr, flags, box);
 935             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
 936         }
 937
 938         /* launch local nonbonded work on GPU */
 939         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
 940         do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::Local, enbvClearFNo,
 941                      step, nrnb, wcycle);
 942         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
 943         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 944     }
 945
 946     if (useGpuPme)
 947     {
 948         // In PME GPU and mixed mode we launch FFT / gather after the
 949         // X copy/transform to allow overlap as well as after the GPU NB
 950         // launch to avoid FFT launch overhead hijacking the CPU and delaying
 951         // the nonbonded kernel.
 952         launchPmeGpuFftAndGather(fr->pmedata, wcycle);
 953     }
 954
 955     /* Communicate coordinates and sum dipole if necessary +
 956        do non-local pair search */
 957     if (havePPDomainDecomposition(cr))
 958     {
 959         if (bNS)
 960         {
 961             wallcycle_start_nocount(wcycle, ewcNS);
 962             wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_NONLOCAL);
 963             /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
 964             nbv->constructPairlist(Nbnxm::InteractionLocality::NonLocal,
 965                                    &top->excls, step, nrnb);
 966             wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_NONLOCAL);
 967             wallcycle_stop(wcycle, ewcNS);
 968         }
 969         else
 970         {
 971             dd_move_x(cr->dd, box, x.unpaddedArrayRef(), wcycle);
 972
 973             nbnxn_atomdata_copy_x_to_nbat_x(nbv->nbs.get(), Nbnxm::AtomLocality::NonLocal,
 974                                             FALSE, as_rvec_array(x.unpaddedArrayRef().data()),
 975                                             nbv->nbat, wcycle);
 976         }
 977
 978         if (bUseGPU)
 979         {
 980             wallcycle_start(wcycle, ewcLAUNCH_GPU);
 981
 982             /* launch non-local nonbonded tasks on GPU */
 983             wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
 984             Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat, Nbnxm::AtomLocality::NonLocal, ppForceWorkload->haveGpuBondedWork);
 985             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
 986
 987             if (ppForceWorkload->haveGpuBondedWork)
 988             {
 989                 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_BONDED);
 990                 fr->gpuBonded->launchKernels(fr, flags, box);
 991                 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
 992             }
 993
 994             wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
 995             do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::NonLocal, enbvClearFNo,
 996                          step, nrnb, wcycle);
 997             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
 998
 999             wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1000         }
1001     }
1002
1003     if (bUseGPU)
1004     {
1005         /* launch D2H copy-back F */
1006         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1007         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1008         if (havePPDomainDecomposition(cr))
1009         {
1010             Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat,
1011                                       flags, Nbnxm::AtomLocality::NonLocal, ppForceWorkload->haveGpuBondedWork);
1012         }
1013         Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat,
1014                                   flags, Nbnxm::AtomLocality::Local, ppForceWorkload->haveGpuBondedWork);
1015         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1016
1017         if (ppForceWorkload->haveGpuBondedWork && (flags & GMX_FORCE_ENERGY))
1018         {
1019             wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_BONDED);
1020             fr->gpuBonded->launchEnergyTransfer();
1021             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1022         }
1023         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1024     }
1025
1026     if (bStateChanged && inputrecNeedMutot(inputrec))
1027     {
1028         if (PAR(cr))
1029         {
1030             gmx_sumd(2*DIM, mu, cr);
1031
1032             ddBalanceRegionHandler.reopenRegionCpu();
1033         }
1034
1035         for (i = 0; i < 2; i++)
1036         {
1037             for (j = 0; j < DIM; j++)
1038             {
1039                 fr->mu_tot[i][j] = mu[i*DIM + j];
1040             }
1041         }
1042     }
1043     if (fr->efep == efepNO)
1044     {
1045         copy_rvec(fr->mu_tot[0], mu_tot);
1046     }
1047     else
1048     {
1049         for (j = 0; j < DIM; j++)
1050         {
1051             mu_tot[j] =
1052                 (1.0 - lambda[efptCOUL])*fr->mu_tot[0][j] +
1053                 lambda[efptCOUL]*fr->mu_tot[1][j];
1054         }
1055     }
1056
1057     /* Reset energies */
1058     reset_enerdata(enerd);
1059     clear_rvecs(SHIFTS, fr->fshift);
1060
1061     if (DOMAINDECOMP(cr) && !thisRankHasDuty(cr, DUTY_PME))
1062     {
1063         wallcycle_start(wcycle, ewcPPDURINGPME);
1064         dd_force_flop_start(cr->dd, nrnb);
1065     }
1066
1067     if (inputrec->bRot)
1068     {
1069         wallcycle_start(wcycle, ewcROT);
1070         do_rotation(cr, enforcedRotation, box, as_rvec_array(x.unpaddedArrayRef().data()), t, step, bNS);
1071         wallcycle_stop(wcycle, ewcROT);
1072     }
1073
1074     /* Temporary solution until all routines take PaddedRVecVector */
1075     rvec *const f = as_rvec_array(force.unpaddedArrayRef().data());
1076
1077     /* Start the force cycle counter.
1078      * Note that a different counter is used for dynamic load balancing.
1079      */
1080     wallcycle_start(wcycle, ewcFORCE);
1081
1082     gmx::ArrayRef<gmx::RVec> forceRef = force.unpaddedArrayRef();
1083     if (bDoForces)
1084     {
1085         /* If we need to compute the virial, we might need a separate
1086          * force buffer for algorithms for which the virial is calculated
1087          * directly, such as PME.
1088          */
1089         if ((flags & GMX_FORCE_VIRIAL) && fr->haveDirectVirialContributions)
1090         {
1091             forceRef = *fr->forceBufferForDirectVirialContributions;
1092
1093             /* TODO: update comment
1094              * We only compute forces on local atoms. Note that vsites can
1095              * spread to non-local atoms, but that part of the buffer is
1096              * cleared separately in the vsite spreading code.
1097              */
1098             clear_rvecs_omp(forceRef.size(), as_rvec_array(forceRef.data()));
1099         }
1100         /* Clear the short- and long-range forces */
1101         clear_rvecs_omp(fr->natoms_force_constr, f);
1102     }
1103
1104     /* forceWithVirial uses the local atom range only */
1105     gmx::ForceWithVirial forceWithVirial(forceRef, (flags & GMX_FORCE_VIRIAL) != 0);
1106
1107     if (inputrec->bPull && pull_have_constraint(inputrec->pull_work))
1108     {
1109         clear_pull_forces(inputrec->pull_work);
1110     }
1111
1112     /* We calculate the non-bonded forces, when done on the CPU, here.
1113      * We do this before calling do_force_lowlevel, because in that
1114      * function, the listed forces are calculated before PME, which
1115      * does communication.  With this order, non-bonded and listed
1116      * force calculation imbalance can be balanced out by the domain
1117      * decomposition load balancing.
1118      */
1119
1120     if (!bUseOrEmulGPU)
1121     {
1122         do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::Local, enbvClearFYes,
1123                      step, nrnb, wcycle);
1124     }
1125
1126     if (fr->efep != efepNO)
1127     {
1128         /* Calculate the local and non-local free energy interactions here.
1129          * Happens here on the CPU both with and without GPU.
1130          */
1131         wallcycle_sub_start(wcycle, ewcsNONBONDED);
1132         nbv->dispatchFreeEnergyKernel(Nbnxm::InteractionLocality::Local,
1133                                       fr, as_rvec_array(x.unpaddedArrayRef().data()), f, *mdatoms,
1134                                       inputrec->fepvals, lambda,
1135                                       enerd, flags, nrnb);
1136
1137         if (havePPDomainDecomposition(cr))
1138         {
1139             nbv->dispatchFreeEnergyKernel(Nbnxm::InteractionLocality::NonLocal,
1140                                           fr, as_rvec_array(x.unpaddedArrayRef().data()), f, *mdatoms,
1141                                           inputrec->fepvals, lambda,
1142                                           enerd, flags, nrnb);
1143         }
1144         wallcycle_sub_stop(wcycle, ewcsNONBONDED);
1145     }
1146
1147     if (!bUseOrEmulGPU)
1148     {
1149         if (havePPDomainDecomposition(cr))
1150         {
1151             do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::NonLocal, enbvClearFNo,
1152                          step, nrnb, wcycle);
1153         }
1154
1155         /* Add all the non-bonded force to the normal force array.
1156          * This can be split into a local and a non-local part when overlapping
1157          * communication with calculation with domain decomposition.
1158          */
1159         wallcycle_stop(wcycle, ewcFORCE);
1160
1161         nbv->atomdata_add_nbat_f_to_f(Nbnxm::AtomLocality::All, f, wcycle);
1162
1163         wallcycle_start_nocount(wcycle, ewcFORCE);
1164
1165         /* If there are multiple fshift output buffers we need to reduce them */
1166         if (flags & GMX_FORCE_VIRIAL)
1167         {
1168             /* This is not in a subcounter because it takes a
1169                negligible and constant-sized amount of time */
1170             nbnxn_atomdata_add_nbat_fshift_to_fshift(nbv->nbat,
1171                                                      fr->fshift);
1172         }
1173     }
1174
1175     /* update QMMMrec, if necessary */
1176     if (fr->bQMMM)
1177     {
1178         update_QMMMrec(cr, fr, as_rvec_array(x.unpaddedArrayRef().data()), mdatoms, box);
1179     }
1180
1181     /* Compute the bonded and non-bonded energies and optionally forces */
1182     do_force_lowlevel(fr, inputrec, &(top->idef),
1183                       cr, ms, nrnb, wcycle, mdatoms,
1184                       as_rvec_array(x.unpaddedArrayRef().data()), hist, f, &forceWithVirial, enerd, fcd,
1185                       box, inputrec->fepvals, lambda, graph, &(top->excls), fr->mu_tot,
1186                       flags,
1187                       &cycles_pme, ddBalanceRegionHandler);
1188
1189     wallcycle_stop(wcycle, ewcFORCE);
1190
1191     computeSpecialForces(fplog, cr, inputrec, awh, enforcedRotation,
1192                          step, t, wcycle,
1193                          fr->forceProviders, box, x.unpaddedArrayRef(), mdatoms, lambda,
1194                          flags, &forceWithVirial, enerd,
1195                          ed, bNS);
1196
1197     if (bUseOrEmulGPU)
1198     {
1199         /* wait for non-local forces (or calculate in emulation mode) */
1200         if (havePPDomainDecomposition(cr))
1201         {
1202             if (bUseGPU)
1203             {
1204                 wallcycle_start(wcycle, ewcWAIT_GPU_NB_NL);
1205                 Nbnxm::gpu_wait_finish_task(nbv->gpu_nbv,
1206                                             flags, Nbnxm::AtomLocality::NonLocal,
1207                                             ppForceWorkload->haveGpuBondedWork,
1208                                             enerd->grpp.ener[egLJSR], enerd->grpp.ener[egCOULSR],
1209                                             fr->fshift);
1210                 cycles_wait_gpu += wallcycle_stop(wcycle, ewcWAIT_GPU_NB_NL);
1211             }
1212             else
1213             {
1214                 wallcycle_start_nocount(wcycle, ewcFORCE);
1215                 do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::NonLocal, enbvClearFYes,
1216                              step, nrnb, wcycle);
1217                 wallcycle_stop(wcycle, ewcFORCE);
1218             }
1219
1220             nbv->atomdata_add_nbat_f_to_f(Nbnxm::AtomLocality::NonLocal,
1221                                           f, wcycle);
1222         }
1223     }
1224
1225     if (havePPDomainDecomposition(cr))
1226     {
1227         /* We are done with the CPU compute.
1228          * We will now communicate the non-local forces.
1229          * If we use a GPU this will overlap with GPU work, so in that case
1230          * we do not close the DD force balancing region here.
1231          */
1232         ddBalanceRegionHandler.closeAfterForceComputationCpu();
1233
1234         if (bDoForces)
1235         {
1236             dd_move_f(cr->dd, force.unpaddedArrayRef(), fr->fshift, wcycle);
1237         }
1238     }
1239
1240     // With both nonbonded and PME offloaded a GPU on the same rank, we use
1241     // an alternating wait/reduction scheme.
1242     bool alternateGpuWait = (!c_disableAlternatingWait && useGpuPme && bUseGPU && !DOMAINDECOMP(cr));
1243     if (alternateGpuWait)
1244     {
1245         alternatePmeNbGpuWaitReduce(fr->nbv, fr->pmedata, &force, &forceWithVirial, fr->fshift, enerd, flags, pmeFlags, ppForceWorkload->haveGpuBondedWork, wcycle);
1246     }
1247
1248     if (!alternateGpuWait && useGpuPme)
1249     {
1250         pme_gpu_wait_and_reduce(fr->pmedata, pmeFlags, wcycle, &forceWithVirial, enerd);
1251     }
1252
1253     /* Wait for local GPU NB outputs on the non-alternating wait path */
1254     if (!alternateGpuWait && bUseGPU)
1255     {
1256         /* Measured overhead on CUDA and OpenCL with(out) GPU sharing
1257          * is between 0.5 and 1.5 Mcycles. So 2 MCycles is an overestimate,
1258          * but even with a step of 0.1 ms the difference is less than 1%
1259          * of the step time.
1260          */
1261         const float gpuWaitApiOverheadMargin = 2e6f; /* cycles */
1262
1263         wallcycle_start(wcycle, ewcWAIT_GPU_NB_L);
1264         Nbnxm::gpu_wait_finish_task(nbv->gpu_nbv,
1265                                     flags, Nbnxm::AtomLocality::Local, ppForceWorkload->haveGpuBondedWork,
1266                                     enerd->grpp.ener[egLJSR], enerd->grpp.ener[egCOULSR],
1267                                     fr->fshift);
1268         float cycles_tmp = wallcycle_stop(wcycle, ewcWAIT_GPU_NB_L);
1269
1270         if (ddBalanceRegionHandler.useBalancingRegion())
1271         {
1272             DdBalanceRegionWaitedForGpu waitedForGpu = DdBalanceRegionWaitedForGpu::yes;
1273             if (bDoForces && cycles_tmp <= gpuWaitApiOverheadMargin)
1274             {
1275                 /* We measured few cycles, it could be that the kernel
1276                  * and transfer finished earlier and there was no actual
1277                  * wait time, only API call overhead.
1278                  * Then the actual time could be anywhere between 0 and
1279                  * cycles_wait_est. We will use half of cycles_wait_est.
1280                  */
1281                 waitedForGpu = DdBalanceRegionWaitedForGpu::no;
1282             }
1283             ddBalanceRegionHandler.closeAfterForceComputationGpu(cycles_wait_gpu, waitedForGpu);
1284         }
1285     }
1286
1287     if (fr->nbv->emulateGpu())
1288     {
1289         // NOTE: emulation kernel is not included in the balancing region,
1290         // but emulation mode does not target performance anyway
1291         wallcycle_start_nocount(wcycle, ewcFORCE);
1292         do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::Local,
1293                      DOMAINDECOMP(cr) ? enbvClearFNo : enbvClearFYes,
1294                      step, nrnb, wcycle);
1295         wallcycle_stop(wcycle, ewcFORCE);
1296     }
1297
1298     if (useGpuPme)
1299     {
1300         pme_gpu_reinit_computation(fr->pmedata, wcycle);
1301     }
1302
1303     if (bUseGPU)
1304     {
1305         /* now clear the GPU outputs while we finish the step on the CPU */
1306         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1307         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1308         Nbnxm::gpu_clear_outputs(nbv->gpu_nbv, flags);
1309
1310         if (nbv->pairlistSets().isDynamicPruningStepGpu(step))
1311         {
1312             nbv->dispatchPruneKernelGpu(step);
1313         }
1314         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1315         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1316     }
1317
1318     if (ppForceWorkload->haveGpuBondedWork && (flags & GMX_FORCE_ENERGY))
1319     {
1320         wallcycle_start(wcycle, ewcWAIT_GPU_BONDED);
1321         // in principle this should be included in the DD balancing region,
1322         // but generally it is infrequent so we'll omit it for the sake of
1323         // simpler code
1324         fr->gpuBonded->accumulateEnergyTerms(enerd);
1325         wallcycle_stop(wcycle, ewcWAIT_GPU_BONDED);
1326
1327         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1328         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_BONDED);
1329         fr->gpuBonded->clearEnergies();
1330         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1331         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1332     }
1333
1334     /* Do the nonbonded GPU (or emulation) force buffer reduction
1335      * on the non-alternating path. */
1336     if (bUseOrEmulGPU && !alternateGpuWait)
1337     {
1338         nbv->atomdata_add_nbat_f_to_f(Nbnxm::AtomLocality::Local,
1339                                       f, wcycle);
1340     }
1341     if (DOMAINDECOMP(cr))
1342     {
1343         dd_force_flop_stop(cr->dd, nrnb);
1344     }
1345
1346     if (bDoForces)
1347     {
1348         /* If we have NoVirSum forces, but we do not calculate the virial,
1349          * we sum fr->f_novirsum=f later.
1350          */
1351         if (vsite && !(fr->haveDirectVirialContributions && !(flags & GMX_FORCE_VIRIAL)))
1352         {
1353             spread_vsite_f(vsite, as_rvec_array(x.unpaddedArrayRef().data()), f, fr->fshift, FALSE, nullptr, nrnb,
1354                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr, wcycle);
1355         }
1356
1357         if (flags & GMX_FORCE_VIRIAL)
1358         {
1359             /* Calculation of the virial must be done after vsites! */
1360             calc_virial(0, mdatoms->homenr, as_rvec_array(x.unpaddedArrayRef().data()), f,
1361                         vir_force, graph, box, nrnb, fr, inputrec->ePBC);
1362         }
1363     }
1364
1365     if (PAR(cr) && !thisRankHasDuty(cr, DUTY_PME))
1366     {
1367         /* In case of node-splitting, the PP nodes receive the long-range
1368          * forces, virial and energy from the PME nodes here.
1369          */
1370         pme_receive_force_ener(cr, &forceWithVirial, enerd, wcycle);
1371     }
1372
1373     if (bDoForces)
1374     {
1375         post_process_forces(cr, step, nrnb, wcycle,
1376                             top, box, as_rvec_array(x.unpaddedArrayRef().data()), f, &forceWithVirial,
1377                             vir_force, mdatoms, graph, fr, vsite,
1378                             flags);
1379     }
1380
1381     if (flags & GMX_FORCE_ENERGY)
1382     {
1383         /* Sum the potential energy terms from group contributions */
1384         sum_epot(&(enerd->grpp), enerd->term);
1385
1386         if (!EI_TPI(inputrec->eI))
1387         {
1388             checkPotentialEnergyValidity(step, *enerd, *inputrec);
1389         }
1390     }
1391 }
1392
1393 static void do_force_cutsGROUP(FILE *fplog,
1394                                const t_commrec *cr,
1395                                const gmx_multisim_t *ms,
1396                                const t_inputrec *inputrec,
1397                                gmx::Awh *awh,
1398                                gmx_enfrot *enforcedRotation,
1399                                int64_t step,
1400                                t_nrnb *nrnb,
1401                                gmx_wallcycle_t wcycle,
1402                                gmx_localtop_t *top,
1403                                const gmx_groups_t *groups,
1404                                matrix box, gmx::ArrayRefWithPadding<gmx::RVec> x,
1405                                history_t *hist,
1406                                gmx::ArrayRefWithPadding<gmx::RVec> force,
1407                                tensor vir_force,
1408                                const t_mdatoms *mdatoms,
1409                                gmx_enerdata_t *enerd,
1410                                t_fcdata *fcd,
1411                                real *lambda,
1412                                t_graph *graph,
1413                                t_forcerec *fr,
1414                                const gmx_vsite_t *vsite,
1415                                rvec mu_tot,
1416                                double t,
1417                                gmx_edsam *ed,
1418                                int flags,
1419                                const DDBalanceRegionHandler &ddBalanceRegionHandler)
1420 {
1421     int        cg0, cg1, i, j;
1422     double     mu[2*DIM];
1423     gmx_bool   bStateChanged, bNS, bFillGrid, bCalcCGCM;
1424     gmx_bool   bDoForces;
1425     float      cycles_pme;
1426
1427     const int  start  = 0;
1428     const int  homenr = mdatoms->homenr;
1429
1430     clear_mat(vir_force);
1431
1432     cg0 = 0;
1433     if (DOMAINDECOMP(cr))
1434     {
1435         cg1 = cr->dd->globalAtomGroupIndices.size();
1436     }
1437     else
1438     {
1439         cg1 = top->cgs.nr;
1440     }
1441     if (fr->n_tpi > 0)
1442     {
1443         cg1--;
1444     }
1445
1446     bStateChanged  = ((flags & GMX_FORCE_STATECHANGED) != 0);
1447     bNS            = ((flags & GMX_FORCE_NS) != 0);
1448     /* Should we perform the long-range nonbonded evaluation inside the neighborsearching? */
1449     bFillGrid      = (bNS && bStateChanged);
1450     bCalcCGCM      = (bFillGrid && !DOMAINDECOMP(cr));
1451     bDoForces      = ((flags & GMX_FORCE_FORCES) != 0);
1452
1453     if (bStateChanged)
1454     {
1455         update_forcerec(fr, box);
1456
1457         if (inputrecNeedMutot(inputrec))
1458         {
1459             /* Calculate total (local) dipole moment in a temporary common array.
1460              * This makes it possible to sum them over nodes faster.
1461              */
1462             calc_mu(start, homenr,
1463                     x.unpaddedArrayRef(), mdatoms->chargeA, mdatoms->chargeB, mdatoms->nChargePerturbed,
1464                     mu, mu+DIM);
1465         }
1466     }
1467
1468     if (fr->ePBC != epbcNONE)
1469     {
1470         /* Compute shift vectors every step,
1471          * because of pressure coupling or box deformation!
1472          */
1473         if ((flags & GMX_FORCE_DYNAMICBOX) && bStateChanged)
1474         {
1475             calc_shifts(box, fr->shift_vec);
1476         }
1477
1478         if (bCalcCGCM)
1479         {
1480             put_charge_groups_in_box(fplog, cg0, cg1, fr->ePBC, box,
1481                                      &(top->cgs), as_rvec_array(x.unpaddedArrayRef().data()), fr->cg_cm);
1482             inc_nrnb(nrnb, eNR_CGCM, homenr);
1483             inc_nrnb(nrnb, eNR_RESETX, cg1-cg0);
1484         }
1485         else if (EI_ENERGY_MINIMIZATION(inputrec->eI) && graph)
1486         {
1487             unshift_self(graph, box, as_rvec_array(x.unpaddedArrayRef().data()));
1488         }
1489     }
1490     else if (bCalcCGCM)
1491     {
1492         calc_cgcm(fplog, cg0, cg1, &(top->cgs), as_rvec_array(x.unpaddedArrayRef().data()), fr->cg_cm);
1493         inc_nrnb(nrnb, eNR_CGCM, homenr);
1494     }
1495
1496     if (bCalcCGCM && gmx_debug_at)
1497     {
1498         pr_rvecs(debug, 0, "cgcm", fr->cg_cm, top->cgs.nr);
1499     }
1500
1501 #if GMX_MPI
1502     if (!thisRankHasDuty(cr, DUTY_PME))
1503     {
1504         /* Send particle coordinates to the pme nodes.
1505          * Since this is only implemented for domain decomposition
1506          * and domain decomposition does not use the graph,
1507          * we do not need to worry about shifting.
1508          */
1509         gmx_pme_send_coordinates(cr, box, as_rvec_array(x.unpaddedArrayRef().data()),
1510                                  lambda[efptCOUL], lambda[efptVDW],
1511                                  (flags & (GMX_FORCE_VIRIAL | GMX_FORCE_ENERGY)) != 0,
1512                                  step, wcycle);
1513     }
1514 #endif /* GMX_MPI */
1515
1516     /* Communicate coordinates and sum dipole if necessary */
1517     if (DOMAINDECOMP(cr))
1518     {
1519         dd_move_x(cr->dd, box, x.unpaddedArrayRef(), wcycle);
1520
1521         /* No GPU support, no move_x overlap, so reopen the balance region here */
1522         ddBalanceRegionHandler.reopenRegionCpu();
1523     }
1524
1525     if (inputrecNeedMutot(inputrec))
1526     {
1527         if (bStateChanged)
1528         {
1529             if (PAR(cr))
1530             {
1531                 gmx_sumd(2*DIM, mu, cr);
1532
1533                 ddBalanceRegionHandler.reopenRegionCpu();
1534             }
1535             for (i = 0; i < 2; i++)
1536             {
1537                 for (j = 0; j < DIM; j++)
1538                 {
1539                     fr->mu_tot[i][j] = mu[i*DIM + j];
1540                 }
1541             }
1542         }
1543         if (fr->efep == efepNO)
1544         {
1545             copy_rvec(fr->mu_tot[0], mu_tot);
1546         }
1547         else
1548         {
1549             for (j = 0; j < DIM; j++)
1550             {
1551                 mu_tot[j] =
1552                     (1.0 - lambda[efptCOUL])*fr->mu_tot[0][j] + lambda[efptCOUL]*fr->mu_tot[1][j];
1553             }
1554         }
1555     }
1556
1557     /* Reset energies */
1558     reset_enerdata(enerd);
1559     clear_rvecs(SHIFTS, fr->fshift);
1560
1561     if (bNS)
1562     {
1563         wallcycle_start(wcycle, ewcNS);
1564
1565         if (graph && bStateChanged)
1566         {
1567             /* Calculate intramolecular shift vectors to make molecules whole */
1568             mk_mshift(fplog, graph, fr->ePBC, box, as_rvec_array(x.unpaddedArrayRef().data()));
1569         }
1570
1571         /* Do the actual neighbour searching */
1572         ns(fplog, fr, box,
1573            groups, top, mdatoms,
1574            cr, nrnb, bFillGrid);
1575
1576         wallcycle_stop(wcycle, ewcNS);
1577     }
1578
1579     if (DOMAINDECOMP(cr) && !thisRankHasDuty(cr, DUTY_PME))
1580     {
1581         wallcycle_start(wcycle, ewcPPDURINGPME);
1582         dd_force_flop_start(cr->dd, nrnb);
1583     }
1584
1585     if (inputrec->bRot)
1586     {
1587         wallcycle_start(wcycle, ewcROT);
1588         do_rotation(cr, enforcedRotation, box, as_rvec_array(x.unpaddedArrayRef().data()), t, step, bNS);
1589         wallcycle_stop(wcycle, ewcROT);
1590     }
1591
1592     /* Temporary solution until all routines take PaddedRVecVector */
1593     rvec *f = as_rvec_array(force.unpaddedArrayRef().data());
1594
1595     /* Start the force cycle counter.
1596      * Note that a different counter is used for dynamic load balancing.
1597      */
1598     wallcycle_start(wcycle, ewcFORCE);
1599
1600     gmx::ArrayRef<gmx::RVec> forceRef = force.paddedArrayRef();
1601     if (bDoForces)
1602     {
1603         /* If we need to compute the virial, we might need a separate
1604          * force buffer for algorithms for which the virial is calculated
1605          * separately, such as PME.
1606          */
1607         if ((flags & GMX_FORCE_VIRIAL) && fr->haveDirectVirialContributions)
1608         {
1609             forceRef = *fr->forceBufferForDirectVirialContributions;
1610
1611             clear_rvecs_omp(forceRef.size(), as_rvec_array(forceRef.data()));
1612         }
1613
1614         /* Clear the short- and long-range forces */
1615         clear_rvecs(fr->natoms_force_constr, f);
1616     }
1617
1618     /* forceWithVirial might need the full force atom range */
1619     gmx::ForceWithVirial forceWithVirial(forceRef, (flags & GMX_FORCE_VIRIAL) != 0);
1620
1621     if (inputrec->bPull && pull_have_constraint(inputrec->pull_work))
1622     {
1623         clear_pull_forces(inputrec->pull_work);
1624     }
1625
1626     /* update QMMMrec, if necessary */
1627     if (fr->bQMMM)
1628     {
1629         update_QMMMrec(cr, fr, as_rvec_array(x.unpaddedArrayRef().data()), mdatoms, box);
1630     }
1631
1632     /* Compute the bonded and non-bonded energies and optionally forces */
1633     do_force_lowlevel(fr, inputrec, &(top->idef),
1634                       cr, ms, nrnb, wcycle, mdatoms,
1635                       as_rvec_array(x.unpaddedArrayRef().data()), hist, f, &forceWithVirial, enerd, fcd,
1636                       box, inputrec->fepvals, lambda,
1637                       graph, &(top->excls), fr->mu_tot,
1638                       flags,
1639                       &cycles_pme, ddBalanceRegionHandler);
1640
1641     wallcycle_stop(wcycle, ewcFORCE);
1642
1643     if (DOMAINDECOMP(cr))
1644     {
1645         dd_force_flop_stop(cr->dd, nrnb);
1646
1647         ddBalanceRegionHandler.closeAfterForceComputationCpu();
1648     }
1649
1650     computeSpecialForces(fplog, cr, inputrec, awh, enforcedRotation,
1651                          step, t, wcycle,
1652                          fr->forceProviders, box, x.unpaddedArrayRef(), mdatoms, lambda,
1653                          flags, &forceWithVirial, enerd,
1654                          ed, bNS);
1655
1656     if (bDoForces)
1657     {
1658         /* Communicate the forces */
1659         if (DOMAINDECOMP(cr))
1660         {
1661             dd_move_f(cr->dd, force.unpaddedArrayRef(), fr->fshift, wcycle);
1662             /* Do we need to communicate the separate force array
1663              * for terms that do not contribute to the single sum virial?
1664              * Position restraints and electric fields do not introduce
1665              * inter-cg forces, only full electrostatics methods do.
1666              * When we do not calculate the virial, fr->f_novirsum = f,
1667              * so we have already communicated these forces.
1668              */
1669             if (EEL_FULL(fr->ic->eeltype) && cr->dd->n_intercg_excl &&
1670                 (flags & GMX_FORCE_VIRIAL))
1671             {
1672                 dd_move_f(cr->dd, forceWithVirial.force_, nullptr, wcycle);
1673             }
1674         }
1675
1676         /* If we have NoVirSum forces, but we do not calculate the virial,
1677          * we sum fr->f_novirsum=f later.
1678          */
1679         if (vsite && !(fr->haveDirectVirialContributions && !(flags & GMX_FORCE_VIRIAL)))
1680         {
1681             spread_vsite_f(vsite, as_rvec_array(x.unpaddedArrayRef().data()), f, fr->fshift, FALSE, nullptr, nrnb,
1682                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr, wcycle);
1683         }
1684
1685         if (flags & GMX_FORCE_VIRIAL)
1686         {
1687             /* Calculation of the virial must be done after vsites! */
1688             calc_virial(0, mdatoms->homenr, as_rvec_array(x.unpaddedArrayRef().data()), f,
1689                         vir_force, graph, box, nrnb, fr, inputrec->ePBC);
1690         }
1691     }
1692
1693     if (PAR(cr) && !thisRankHasDuty(cr, DUTY_PME))
1694     {
1695         /* In case of node-splitting, the PP nodes receive the long-range
1696          * forces, virial and energy from the PME nodes here.
1697          */
1698         pme_receive_force_ener(cr, &forceWithVirial, enerd, wcycle);
1699     }
1700
1701     if (bDoForces)
1702     {
1703         post_process_forces(cr, step, nrnb, wcycle,
1704                             top, box, as_rvec_array(x.unpaddedArrayRef().data()), f, &forceWithVirial,
1705                             vir_force, mdatoms, graph, fr, vsite,
1706                             flags);
1707     }
1708
1709     if (flags & GMX_FORCE_ENERGY)
1710     {
1711         /* Sum the potential energy terms from group contributions */
1712         sum_epot(&(enerd->grpp), enerd->term);
1713
1714         if (!EI_TPI(inputrec->eI))
1715         {
1716             checkPotentialEnergyValidity(step, *enerd, *inputrec);
1717         }
1718     }
1719
1720 }
1721
1722 void do_force(FILE                                     *fplog,
1723               const t_commrec                          *cr,
1724               const gmx_multisim_t                     *ms,
1725               const t_inputrec                         *inputrec,
1726               gmx::Awh                                 *awh,
1727               gmx_enfrot                               *enforcedRotation,
1728               int64_t                                   step,
1729               t_nrnb                                   *nrnb,
1730               gmx_wallcycle_t                           wcycle,
1731               gmx_localtop_t                           *top,
1732               const gmx_groups_t                       *groups,
1733               matrix                                    box,
1734               gmx::ArrayRefWithPadding<gmx::RVec>       x,     //NOLINT(performance-unnecessary-value-param)
1735               history_t                                *hist,
1736               gmx::ArrayRefWithPadding<gmx::RVec>       force, //NOLINT(performance-unnecessary-value-param)
1737               tensor                                    vir_force,
1738               const t_mdatoms                          *mdatoms,
1739               gmx_enerdata_t                           *enerd,
1740               t_fcdata                                 *fcd,
1741               gmx::ArrayRef<real>                       lambda,
1742               t_graph                                  *graph,
1743               t_forcerec                               *fr,
1744               gmx::PpForceWorkload                     *ppForceWorkload,
1745               const gmx_vsite_t                        *vsite,
1746               rvec                                      mu_tot,
1747               double                                    t,
1748               gmx_edsam                                *ed,
1749               int                                       flags,
1750               const DDBalanceRegionHandler             &ddBalanceRegionHandler)
1751 {
1752     /* modify force flag if not doing nonbonded */
1753     if (!fr->bNonbonded)
1754     {
1755         flags &= ~GMX_FORCE_NONBONDED;
1756     }
1757
1758     switch (inputrec->cutoff_scheme)
1759     {
1760         case ecutsVERLET:
1761             do_force_cutsVERLET(fplog, cr, ms, inputrec,
1762                                 awh, enforcedRotation, step, nrnb, wcycle,
1763                                 top,
1764                                 groups,
1765                                 box, x, hist,
1766                                 force, vir_force,
1767                                 mdatoms,
1768                                 enerd, fcd,
1769                                 lambda.data(), graph,
1770                                 fr,
1771                                 ppForceWorkload,
1772                                 fr->ic,
1773                                 vsite, mu_tot,
1774                                 t, ed,
1775                                 flags,
1776                                 ddBalanceRegionHandler);
1777             break;
1778         case ecutsGROUP:
1779             do_force_cutsGROUP(fplog, cr, ms, inputrec,
1780                                awh, enforcedRotation, step, nrnb, wcycle,
1781                                top,
1782                                groups,
1783                                box, x, hist,
1784                                force, vir_force,
1785                                mdatoms,
1786                                enerd, fcd,
1787                                lambda.data(), graph,
1788                                fr, vsite, mu_tot,
1789                                t, ed,
1790                                flags,
1791                                ddBalanceRegionHandler);
1792             break;
1793         default:
1794             gmx_incons("Invalid cut-off scheme passed!");
1795     }
1796
1797     /* In case we don't have constraints and are using GPUs, the next balancing
1798      * region starts here.
1799      * Some "special" work at the end of do_force_cuts?, such as vsite spread,
1800      * virial calculation and COM pulling, is not thus not included in
1801      * the balance timing, which is ok as most tasks do communication.
1802      */
1803     ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::no);
1804 }
1805
1806
1807 void do_constrain_first(FILE *fplog, gmx::Constraints *constr,
1808                         const t_inputrec *ir, const t_mdatoms *md,
1809                         t_state *state)
1810 {
1811     int             i, m, start, end;
1812     int64_t         step;
1813     real            dt = ir->delta_t;
1814     real            dvdl_dum;
1815     rvec           *savex;
1816
1817     /* We need to allocate one element extra, since we might use
1818      * (unaligned) 4-wide SIMD loads to access rvec entries.
1819      */
1820     snew(savex, state->natoms + 1);
1821
1822     start = 0;
1823     end   = md->homenr;
1824
1825     if (debug)
1826     {
1827         fprintf(debug, "vcm: start=%d, homenr=%d, end=%d\n",
1828                 start, md->homenr, end);
1829     }
1830     /* Do a first constrain to reset particles... */
1831     step = ir->init_step;
1832     if (fplog)
1833     {
1834         char buf[STEPSTRSIZE];
1835         fprintf(fplog, "\nConstraining the starting coordinates (step %s)\n",
1836                 gmx_step_str(step, buf));
1837     }
1838     dvdl_dum = 0;
1839
1840     /* constrain the current position */
1841     constr->apply(TRUE, FALSE,
1842                   step, 0, 1.0,
1843                   state->x.rvec_array(), state->x.rvec_array(), nullptr,
1844                   state->box,
1845                   state->lambda[efptBONDED], &dvdl_dum,
1846                   nullptr, nullptr, gmx::ConstraintVariable::Positions);
1847     if (EI_VV(ir->eI))
1848     {
1849         /* constrain the inital velocity, and save it */
1850         /* also may be useful if we need the ekin from the halfstep for velocity verlet */
1851         constr->apply(TRUE, FALSE,
1852                       step, 0, 1.0,
1853                       state->x.rvec_array(), state->v.rvec_array(), state->v.rvec_array(),
1854                       state->box,
1855                       state->lambda[efptBONDED], &dvdl_dum,
1856                       nullptr, nullptr, gmx::ConstraintVariable::Velocities);
1857     }
1858     /* constrain the inital velocities at t-dt/2 */
1859     if (EI_STATE_VELOCITY(ir->eI) && ir->eI != eiVV)
1860     {
1861         auto x = makeArrayRef(state->x).subArray(start, end);
1862         auto v = makeArrayRef(state->v).subArray(start, end);
1863         for (i = start; (i < end); i++)
1864         {
1865             for (m = 0; (m < DIM); m++)
1866             {
1867                 /* Reverse the velocity */
1868                 v[i][m] = -v[i][m];
1869                 /* Store the position at t-dt in buf */
1870                 savex[i][m] = x[i][m] + dt*v[i][m];
1871             }
1872         }
1873         /* Shake the positions at t=-dt with the positions at t=0
1874          * as reference coordinates.
1875          */
1876         if (fplog)
1877         {
1878             char buf[STEPSTRSIZE];
1879             fprintf(fplog, "\nConstraining the coordinates at t0-dt (step %s)\n",
1880                     gmx_step_str(step, buf));
1881         }
1882         dvdl_dum = 0;
1883         constr->apply(TRUE, FALSE,
1884                       step, -1, 1.0,
1885                       state->x.rvec_array(), savex, nullptr,
1886                       state->box,
1887                       state->lambda[efptBONDED], &dvdl_dum,
1888                       state->v.rvec_array(), nullptr, gmx::ConstraintVariable::Positions);
1889
1890         for (i = start; i < end; i++)
1891         {
1892             for (m = 0; m < DIM; m++)
1893             {
1894                 /* Re-reverse the velocities */
1895                 v[i][m] = -v[i][m];
1896             }
1897         }
1898     }
1899     sfree(savex);
1900 }
1901
1902 void put_atoms_in_box_omp(int ePBC, const matrix box, gmx::ArrayRef<gmx::RVec> x)
1903 {
1904     int t, nth;
1905     nth = gmx_omp_nthreads_get(emntDefault);
1906
1907 #pragma omp parallel for num_threads(nth) schedule(static)
1908     for (t = 0; t < nth; t++)
1909     {
1910         try
1911         {
1912             size_t natoms = x.size();
1913             size_t offset = (natoms*t    )/nth;
1914             size_t len    = (natoms*(t + 1))/nth - offset;
1915             put_atoms_in_box(ePBC, box, x.subArray(offset, len));
1916         }
1917         GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1918     }
1919 }
1920
1921 void initialize_lambdas(FILE               *fplog,
1922                         const t_inputrec   &ir,
1923                         bool                isMaster,
1924                         int                *fep_state,
1925                         gmx::ArrayRef<real> lambda,
1926                         double             *lam0)
1927 {
1928     /* TODO: Clean up initialization of fep_state and lambda in
1929        t_state.  This function works, but could probably use a logic
1930        rewrite to keep all the different types of efep straight. */
1931
1932     if ((ir.efep == efepNO) && (!ir.bSimTemp))
1933     {
1934         return;
1935     }
1936
1937     const t_lambda *fep = ir.fepvals;
1938     if (isMaster)
1939     {
1940         *fep_state = fep->init_fep_state; /* this might overwrite the checkpoint
1941                                              if checkpoint is set -- a kludge is in for now
1942                                              to prevent this.*/
1943     }
1944
1945     for (int i = 0; i < efptNR; i++)
1946     {
1947         double thisLambda;
1948         /* overwrite lambda state with init_lambda for now for backwards compatibility */
1949         if (fep->init_lambda >= 0) /* if it's -1, it was never initialized */
1950         {
1951             thisLambda = fep->init_lambda;
1952         }
1953         else
1954         {
1955             thisLambda = fep->all_lambda[i][fep->init_fep_state];
1956         }
1957         if (isMaster)
1958         {
1959             lambda[i] = thisLambda;
1960         }
1961         if (lam0 != nullptr)
1962         {
1963             lam0[i] = thisLambda;
1964         }
1965     }
1966     if (ir.bSimTemp)
1967     {
1968         /* need to rescale control temperatures to match current state */
1969         for (int i = 0; i < ir.opts.ngtc; i++)
1970         {
1971             if (ir.opts.ref_t[i] > 0)
1972             {
1973                 ir.opts.ref_t[i] = ir.simtempvals->temperatures[fep->init_fep_state];
1974             }
1975         }
1976     }
1977
1978     /* Send to the log the information on the current lambdas */
1979     if (fplog != nullptr)
1980     {
1981         fprintf(fplog, "Initial vector of lambda components:[ ");
1982         for (const auto &l : lambda)
1983         {
1984             fprintf(fplog, "%10.4f ", l);
1985         }
1986         fprintf(fplog, "]\n");
1987     }
1988 }