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.get();
 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 /*! \brief Hack structure with force ouput buffers for do_force */
 686 struct ForceOutputs
 687 {
 688     //! Constructor
 689     ForceOutputs(rvec *f, gmx::ForceWithVirial const forceWithVirial) :
 690         f(f),
 691         forceWithVirial(forceWithVirial) {}
 692
 693     //! Force output buffer used by legacy modules
 694     rvec                 *const f;
 695     //! Force with direct virial contribution (if there are any)
 696     gmx::ForceWithVirial        forceWithVirial;
 697 };
 698
 699 /*! \brief Set up the different force buffers; also does clearing.
 700  *
 701  * \param[in] fr        force record pointer
 702  * \param[in] inputrec  input record
 703  * \param[in] force     force array
 704  * \param[in] bDoForces True if force are computed this step
 705  * \param[in] doVirial  True if virial is computed this step
 706  * \param[out] wcycle   wallcycle recording structure
 707  *
 708  * \returns             Cleared force output structure
 709  */
 710 static ForceOutputs
 711 setupForceOutputs(const t_forcerec                    *fr,
 712                   const t_inputrec                    &inputrec,
 713                   gmx::ArrayRefWithPadding<gmx::RVec>  force,
 714                   const bool                           bDoForces,
 715                   const bool                           doVirial,
 716                   gmx_wallcycle_t                      wcycle)
 717 {
 718     wallcycle_sub_start(wcycle, ewcsCLEAR_FORCE_BUFFER);
 719
 720     /* Temporary solution until all routines take PaddedRVecVector */
 721     rvec *const f = as_rvec_array(force.unpaddedArrayRef().data());
 722     if (bDoForces)
 723     {
 724         /* Clear the short- and long-range forces */
 725         clear_rvecs_omp(fr->natoms_force_constr, f);
 726     }
 727
 728     /* If we need to compute the virial, we might need a separate
 729      * force buffer for algorithms for which the virial is calculated
 730      * directly, such as PME. Otherwise, forceWithVirial uses the
 731      * the same force (f in legacy calls) buffer as other algorithms.
 732      */
 733     const bool useSeparateForceWithVirialBuffer = (bDoForces && (doVirial && fr->haveDirectVirialContributions));
 734
 735
 736     /* forceWithVirial uses the local atom range only */
 737     gmx::ForceWithVirial forceWithVirial (useSeparateForceWithVirialBuffer ?
 738                                           *fr->forceBufferForDirectVirialContributions : force.unpaddedArrayRef(),
 739                                           doVirial);
 740
 741     if (useSeparateForceWithVirialBuffer)
 742     {
 743         /* TODO: update comment
 744          * We only compute forces on local atoms. Note that vsites can
 745          * spread to non-local atoms, but that part of the buffer is
 746          * cleared separately in the vsite spreading code.
 747          */
 748         clear_rvecs_omp(forceWithVirial.force_.size(), as_rvec_array(forceWithVirial.force_.data()));
 749     }
 750
 751     if (inputrec.bPull && pull_have_constraint(inputrec.pull_work))
 752     {
 753         clear_pull_forces(inputrec.pull_work);
 754     }
 755
 756     wallcycle_sub_stop(wcycle, ewcsCLEAR_FORCE_BUFFER);
 757
 758     return ForceOutputs(f, forceWithVirial);
 759 }
 760
 761
 762 static void do_force_cutsVERLET(FILE *fplog,
 763                                 const t_commrec *cr,
 764                                 const gmx_multisim_t *ms,
 765                                 const t_inputrec *inputrec,
 766                                 gmx::Awh *awh,
 767                                 gmx_enfrot *enforcedRotation,
 768                                 int64_t step,
 769                                 t_nrnb *nrnb,
 770                                 gmx_wallcycle_t wcycle,
 771                                 const gmx_localtop_t *top,
 772                                 const gmx_groups_t * /* groups */,
 773                                 matrix box, gmx::ArrayRefWithPadding<gmx::RVec> x,
 774                                 history_t *hist,
 775                                 gmx::ArrayRefWithPadding<gmx::RVec> force,
 776                                 tensor vir_force,
 777                                 const t_mdatoms *mdatoms,
 778                                 gmx_enerdata_t *enerd, t_fcdata *fcd,
 779                                 real *lambda,
 780                                 t_graph *graph,
 781                                 t_forcerec *fr,
 782                                 gmx::PpForceWorkload *ppForceWorkload,
 783                                 interaction_const_t *ic,
 784                                 const gmx_vsite_t *vsite,
 785                                 rvec mu_tot,
 786                                 double t,
 787                                 gmx_edsam *ed,
 788                                 const int flags,
 789                                 const DDBalanceRegionHandler &ddBalanceRegionHandler)
 790 {
 791     int                 cg1, i, j;
 792     double              mu[2*DIM];
 793     gmx_bool            bStateChanged, bNS, bFillGrid, bCalcCGCM;
 794     gmx_bool            bDoForces, bUseGPU, bUseOrEmulGPU;
 795     rvec                vzero, box_diag;
 796     float               cycles_pme, cycles_wait_gpu;
 797     nonbonded_verlet_t *nbv = fr->nbv.get();
 798
 799     bStateChanged = ((flags & GMX_FORCE_STATECHANGED) != 0);
 800     bNS           = ((flags & GMX_FORCE_NS) != 0);
 801     bFillGrid     = (bNS && bStateChanged);
 802     bCalcCGCM     = (bFillGrid && !DOMAINDECOMP(cr));
 803     bDoForces     = ((flags & GMX_FORCE_FORCES) != 0);
 804     bUseGPU       = fr->nbv->useGpu();
 805     bUseOrEmulGPU = bUseGPU || fr->nbv->emulateGpu();
 806
 807     const auto pmeRunMode = fr->pmedata ? pme_run_mode(fr->pmedata) : PmeRunMode::CPU;
 808     // TODO slim this conditional down - inputrec and duty checks should mean the same in proper code!
 809     const bool useGpuPme  = EEL_PME(fr->ic->eeltype) && thisRankHasDuty(cr, DUTY_PME) &&
 810         ((pmeRunMode == PmeRunMode::GPU) || (pmeRunMode == PmeRunMode::Mixed));
 811     const int  pmeFlags = GMX_PME_SPREAD | GMX_PME_SOLVE |
 812         ((flags & GMX_FORCE_VIRIAL) ? GMX_PME_CALC_ENER_VIR : 0) |
 813         ((flags & GMX_FORCE_ENERGY) ? GMX_PME_CALC_ENER_VIR : 0) |
 814         ((flags & GMX_FORCE_FORCES) ? GMX_PME_CALC_F : 0);
 815
 816     /* At a search step we need to start the first balancing region
 817      * somewhere early inside the step after communication during domain
 818      * decomposition (and not during the previous step as usual).
 819      */
 820     if (bNS)
 821     {
 822         ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::yes);
 823     }
 824
 825     cycles_wait_gpu = 0;
 826
 827     const int start  = 0;
 828     const int homenr = mdatoms->homenr;
 829
 830     clear_mat(vir_force);
 831
 832     if (DOMAINDECOMP(cr))
 833     {
 834         cg1 = cr->dd->globalAtomGroupIndices.size();
 835     }
 836     else
 837     {
 838         cg1 = top->cgs.nr;
 839     }
 840     if (fr->n_tpi > 0)
 841     {
 842         cg1--;
 843     }
 844
 845     if (bStateChanged)
 846     {
 847         update_forcerec(fr, box);
 848
 849         if (inputrecNeedMutot(inputrec))
 850         {
 851             /* Calculate total (local) dipole moment in a temporary common array.
 852              * This makes it possible to sum them over nodes faster.
 853              */
 854             calc_mu(start, homenr,
 855                     x.unpaddedArrayRef(), mdatoms->chargeA, mdatoms->chargeB, mdatoms->nChargePerturbed,
 856                     mu, mu+DIM);
 857         }
 858     }
 859
 860     if (fr->ePBC != epbcNONE)
 861     {
 862         /* Compute shift vectors every step,
 863          * because of pressure coupling or box deformation!
 864          */
 865         if ((flags & GMX_FORCE_DYNAMICBOX) && bStateChanged)
 866         {
 867             calc_shifts(box, fr->shift_vec);
 868         }
 869
 870         if (bCalcCGCM)
 871         {
 872             put_atoms_in_box_omp(fr->ePBC, box, x.unpaddedArrayRef().subArray(0, homenr));
 873             inc_nrnb(nrnb, eNR_SHIFTX, homenr);
 874         }
 875         else if (EI_ENERGY_MINIMIZATION(inputrec->eI) && graph)
 876         {
 877             unshift_self(graph, box, as_rvec_array(x.unpaddedArrayRef().data()));
 878         }
 879     }
 880
 881     nbnxn_atomdata_copy_shiftvec((flags & GMX_FORCE_DYNAMICBOX) != 0,
 882                                  fr->shift_vec, nbv->nbat.get());
 883
 884 #if GMX_MPI
 885     if (!thisRankHasDuty(cr, DUTY_PME))
 886     {
 887         /* Send particle coordinates to the pme nodes.
 888          * Since this is only implemented for domain decomposition
 889          * and domain decomposition does not use the graph,
 890          * we do not need to worry about shifting.
 891          */
 892         gmx_pme_send_coordinates(cr, box, as_rvec_array(x.unpaddedArrayRef().data()),
 893                                  lambda[efptCOUL], lambda[efptVDW],
 894                                  (flags & (GMX_FORCE_VIRIAL | GMX_FORCE_ENERGY)) != 0,
 895                                  step, wcycle);
 896     }
 897 #endif /* GMX_MPI */
 898
 899     if (useGpuPme)
 900     {
 901         launchPmeGpuSpread(fr->pmedata, box, as_rvec_array(x.unpaddedArrayRef().data()), flags, pmeFlags, wcycle);
 902     }
 903
 904     /* do gridding for pair search */
 905     if (bNS)
 906     {
 907         if (graph && bStateChanged)
 908         {
 909             /* Calculate intramolecular shift vectors to make molecules whole */
 910             mk_mshift(fplog, graph, fr->ePBC, box, as_rvec_array(x.unpaddedArrayRef().data()));
 911         }
 912
 913         clear_rvec(vzero);
 914         box_diag[XX] = box[XX][XX];
 915         box_diag[YY] = box[YY][YY];
 916         box_diag[ZZ] = box[ZZ][ZZ];
 917
 918         wallcycle_start(wcycle, ewcNS);
 919         if (!DOMAINDECOMP(cr))
 920         {
 921             wallcycle_sub_start(wcycle, ewcsNBS_GRID_LOCAL);
 922             nbnxn_put_on_grid(nbv, box,
 923                               0, vzero, box_diag,
 924                               nullptr, 0, mdatoms->homenr, -1,
 925                               fr->cginfo, x.unpaddedArrayRef(),
 926                               0, nullptr);
 927             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_LOCAL);
 928         }
 929         else
 930         {
 931             wallcycle_sub_start(wcycle, ewcsNBS_GRID_NONLOCAL);
 932             nbnxn_put_on_grid_nonlocal(nbv, domdec_zones(cr->dd),
 933                                        fr->cginfo, x.unpaddedArrayRef());
 934             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_NONLOCAL);
 935         }
 936
 937         nbnxn_atomdata_set(nbv->nbat.get(), nbv->nbs.get(), mdatoms, fr->cginfo);
 938
 939         wallcycle_stop(wcycle, ewcNS);
 940     }
 941
 942     /* initialize the GPU atom data and copy shift vector */
 943     if (bUseGPU)
 944     {
 945         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 946         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
 947
 948         if (bNS)
 949         {
 950             Nbnxm::gpu_init_atomdata(nbv->gpu_nbv, nbv->nbat.get());
 951         }
 952
 953         Nbnxm::gpu_upload_shiftvec(nbv->gpu_nbv, nbv->nbat.get());
 954
 955         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
 956
 957         if (bNS && fr->gpuBonded)
 958         {
 959             /* Now we put all atoms on the grid, we can assign bonded
 960              * interactions to the GPU, where the grid order is
 961              * needed. Also the xq, f and fshift device buffers have
 962              * been reallocated if needed, so the bonded code can
 963              * learn about them. */
 964             // TODO the xq, f, and fshift buffers are now shared
 965             // resources, so they should be maintained by a
 966             // higher-level object than the nb module.
 967             fr->gpuBonded->updateInteractionListsAndDeviceBuffers(nbnxn_get_gridindices(fr->nbv->nbs.get()),
 968                                                                   top->idef,
 969                                                                   Nbnxm::gpu_get_xq(nbv->gpu_nbv),
 970                                                                   Nbnxm::gpu_get_f(nbv->gpu_nbv),
 971                                                                   Nbnxm::gpu_get_fshift(nbv->gpu_nbv));
 972             ppForceWorkload->haveGpuBondedWork = fr->gpuBonded->haveInteractions();
 973         }
 974
 975         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 976     }
 977
 978     /* do local pair search */
 979     if (bNS)
 980     {
 981         wallcycle_start_nocount(wcycle, ewcNS);
 982         wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_LOCAL);
 983         /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
 984         nbv->constructPairlist(Nbnxm::InteractionLocality::Local,
 985                                &top->excls, step, nrnb);
 986         wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_LOCAL);
 987         wallcycle_stop(wcycle, ewcNS);
 988     }
 989     else
 990     {
 991         nbnxn_atomdata_copy_x_to_nbat_x(nbv->nbs.get(), Nbnxm::AtomLocality::Local,
 992                                         FALSE, as_rvec_array(x.unpaddedArrayRef().data()),
 993                                         nbv->nbat.get(), wcycle);
 994     }
 995
 996     if (bUseGPU)
 997     {
 998         ddBalanceRegionHandler.openBeforeForceComputationGpu();
 999
1000         wallcycle_start(wcycle, ewcLAUNCH_GPU);
1001
1002         wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1003         Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(),
1004                                   Nbnxm::AtomLocality::Local,
1005                                   ppForceWorkload->haveGpuBondedWork);
1006         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1007
1008         // bonded work not split into separate local and non-local, so with DD
1009         // we can only launch the kernel after non-local coordinates have been received.
1010         if (ppForceWorkload->haveGpuBondedWork && !havePPDomainDecomposition(cr))
1011         {
1012             wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_BONDED);
1013             fr->gpuBonded->launchKernels(fr, flags, box);
1014             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1015         }
1016
1017         /* launch local nonbonded work on GPU */
1018         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1019         do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::Local, enbvClearFNo,
1020                      step, nrnb, wcycle);
1021         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1022         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1023     }
1024
1025     if (useGpuPme)
1026     {
1027         // In PME GPU and mixed mode we launch FFT / gather after the
1028         // X copy/transform to allow overlap as well as after the GPU NB
1029         // launch to avoid FFT launch overhead hijacking the CPU and delaying
1030         // the nonbonded kernel.
1031         launchPmeGpuFftAndGather(fr->pmedata, wcycle);
1032     }
1033
1034     /* Communicate coordinates and sum dipole if necessary +
1035        do non-local pair search */
1036     if (havePPDomainDecomposition(cr))
1037     {
1038         if (bNS)
1039         {
1040             wallcycle_start_nocount(wcycle, ewcNS);
1041             wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1042             /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
1043             nbv->constructPairlist(Nbnxm::InteractionLocality::NonLocal,
1044                                    &top->excls, step, nrnb);
1045             wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1046             wallcycle_stop(wcycle, ewcNS);
1047         }
1048         else
1049         {
1050             dd_move_x(cr->dd, box, x.unpaddedArrayRef(), wcycle);
1051
1052             nbnxn_atomdata_copy_x_to_nbat_x(nbv->nbs.get(), Nbnxm::AtomLocality::NonLocal,
1053                                             FALSE, as_rvec_array(x.unpaddedArrayRef().data()),
1054                                             nbv->nbat.get(), wcycle);
1055         }
1056
1057         if (bUseGPU)
1058         {
1059             wallcycle_start(wcycle, ewcLAUNCH_GPU);
1060
1061             /* launch non-local nonbonded tasks on GPU */
1062             wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1063             Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(),
1064                                       Nbnxm::AtomLocality::NonLocal,
1065                                       ppForceWorkload->haveGpuBondedWork);
1066             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1067
1068             if (ppForceWorkload->haveGpuBondedWork)
1069             {
1070                 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_BONDED);
1071                 fr->gpuBonded->launchKernels(fr, flags, box);
1072                 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1073             }
1074
1075             wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1076             do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::NonLocal, enbvClearFNo,
1077                          step, nrnb, wcycle);
1078             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1079
1080             wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1081         }
1082     }
1083
1084     if (bUseGPU)
1085     {
1086         /* launch D2H copy-back F */
1087         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1088         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1089         if (havePPDomainDecomposition(cr))
1090         {
1091             Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat.get(),
1092                                       flags, Nbnxm::AtomLocality::NonLocal, ppForceWorkload->haveGpuBondedWork);
1093         }
1094         Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat.get(),
1095                                   flags, Nbnxm::AtomLocality::Local, ppForceWorkload->haveGpuBondedWork);
1096         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1097
1098         if (ppForceWorkload->haveGpuBondedWork && (flags & GMX_FORCE_ENERGY))
1099         {
1100             wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_BONDED);
1101             fr->gpuBonded->launchEnergyTransfer();
1102             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1103         }
1104         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1105     }
1106
1107     if (bStateChanged && inputrecNeedMutot(inputrec))
1108     {
1109         if (PAR(cr))
1110         {
1111             gmx_sumd(2*DIM, mu, cr);
1112
1113             ddBalanceRegionHandler.reopenRegionCpu();
1114         }
1115
1116         for (i = 0; i < 2; i++)
1117         {
1118             for (j = 0; j < DIM; j++)
1119             {
1120                 fr->mu_tot[i][j] = mu[i*DIM + j];
1121             }
1122         }
1123     }
1124     if (fr->efep == efepNO)
1125     {
1126         copy_rvec(fr->mu_tot[0], mu_tot);
1127     }
1128     else
1129     {
1130         for (j = 0; j < DIM; j++)
1131         {
1132             mu_tot[j] =
1133                 (1.0 - lambda[efptCOUL])*fr->mu_tot[0][j] +
1134                 lambda[efptCOUL]*fr->mu_tot[1][j];
1135         }
1136     }
1137
1138     /* Reset energies */
1139     reset_enerdata(enerd);
1140     clear_rvecs(SHIFTS, fr->fshift);
1141
1142     if (DOMAINDECOMP(cr) && !thisRankHasDuty(cr, DUTY_PME))
1143     {
1144         wallcycle_start(wcycle, ewcPPDURINGPME);
1145         dd_force_flop_start(cr->dd, nrnb);
1146     }
1147
1148     if (inputrec->bRot)
1149     {
1150         wallcycle_start(wcycle, ewcROT);
1151         do_rotation(cr, enforcedRotation, box, as_rvec_array(x.unpaddedArrayRef().data()), t, step, bNS);
1152         wallcycle_stop(wcycle, ewcROT);
1153     }
1154
1155     /* Start the force cycle counter.
1156      * Note that a different counter is used for dynamic load balancing.
1157      */
1158     wallcycle_start(wcycle, ewcFORCE);
1159
1160     // set up and clear force outputs
1161     struct ForceOutputs forceOut = setupForceOutputs(fr, *inputrec, force, bDoForces, ((flags & GMX_FORCE_VIRIAL) != 0), wcycle);
1162
1163     /* We calculate the non-bonded forces, when done on the CPU, here.
1164      * We do this before calling do_force_lowlevel, because in that
1165      * function, the listed forces are calculated before PME, which
1166      * does communication.  With this order, non-bonded and listed
1167      * force calculation imbalance can be balanced out by the domain
1168      * decomposition load balancing.
1169      */
1170
1171     if (!bUseOrEmulGPU)
1172     {
1173         do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::Local, enbvClearFYes,
1174                      step, nrnb, wcycle);
1175     }
1176
1177     if (fr->efep != efepNO)
1178     {
1179         /* Calculate the local and non-local free energy interactions here.
1180          * Happens here on the CPU both with and without GPU.
1181          */
1182         wallcycle_sub_start(wcycle, ewcsNONBONDED);
1183         nbv->dispatchFreeEnergyKernel(Nbnxm::InteractionLocality::Local,
1184                                       fr, as_rvec_array(x.unpaddedArrayRef().data()), forceOut.f, *mdatoms,
1185                                       inputrec->fepvals, lambda,
1186                                       enerd, flags, nrnb);
1187
1188         if (havePPDomainDecomposition(cr))
1189         {
1190             nbv->dispatchFreeEnergyKernel(Nbnxm::InteractionLocality::NonLocal,
1191                                           fr, as_rvec_array(x.unpaddedArrayRef().data()), forceOut.f, *mdatoms,
1192                                           inputrec->fepvals, lambda,
1193                                           enerd, flags, nrnb);
1194         }
1195         wallcycle_sub_stop(wcycle, ewcsNONBONDED);
1196     }
1197
1198     if (!bUseOrEmulGPU)
1199     {
1200         if (havePPDomainDecomposition(cr))
1201         {
1202             do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::NonLocal, enbvClearFNo,
1203                          step, nrnb, wcycle);
1204         }
1205
1206         /* Add all the non-bonded force to the normal force array.
1207          * This can be split into a local and a non-local part when overlapping
1208          * communication with calculation with domain decomposition.
1209          */
1210         wallcycle_stop(wcycle, ewcFORCE);
1211
1212         nbv->atomdata_add_nbat_f_to_f(Nbnxm::AtomLocality::All, forceOut.f, wcycle);
1213
1214         wallcycle_start_nocount(wcycle, ewcFORCE);
1215
1216         /* If there are multiple fshift output buffers we need to reduce them */
1217         if (flags & GMX_FORCE_VIRIAL)
1218         {
1219             /* This is not in a subcounter because it takes a
1220                negligible and constant-sized amount of time */
1221             nbnxn_atomdata_add_nbat_fshift_to_fshift(nbv->nbat.get(),
1222                                                      fr->fshift);
1223         }
1224     }
1225
1226     /* update QMMMrec, if necessary */
1227     if (fr->bQMMM)
1228     {
1229         update_QMMMrec(cr, fr, as_rvec_array(x.unpaddedArrayRef().data()), mdatoms, box);
1230     }
1231
1232     /* Compute the bonded and non-bonded energies and optionally forces */
1233     do_force_lowlevel(fr, inputrec, &(top->idef),
1234                       cr, ms, nrnb, wcycle, mdatoms,
1235                       as_rvec_array(x.unpaddedArrayRef().data()), hist, forceOut.f, &forceOut.forceWithVirial, enerd, fcd,
1236                       box, inputrec->fepvals, lambda, graph, &(top->excls), fr->mu_tot,
1237                       flags,
1238                       &cycles_pme, ddBalanceRegionHandler);
1239
1240     wallcycle_stop(wcycle, ewcFORCE);
1241
1242     computeSpecialForces(fplog, cr, inputrec, awh, enforcedRotation,
1243                          step, t, wcycle,
1244                          fr->forceProviders, box, x.unpaddedArrayRef(), mdatoms, lambda,
1245                          flags, &forceOut.forceWithVirial, enerd,
1246                          ed, bNS);
1247
1248     if (bUseOrEmulGPU)
1249     {
1250         /* wait for non-local forces (or calculate in emulation mode) */
1251         if (havePPDomainDecomposition(cr))
1252         {
1253             if (bUseGPU)
1254             {
1255                 wallcycle_start(wcycle, ewcWAIT_GPU_NB_NL);
1256                 Nbnxm::gpu_wait_finish_task(nbv->gpu_nbv,
1257                                             flags, Nbnxm::AtomLocality::NonLocal,
1258                                             ppForceWorkload->haveGpuBondedWork,
1259                                             enerd->grpp.ener[egLJSR], enerd->grpp.ener[egCOULSR],
1260                                             fr->fshift);
1261                 cycles_wait_gpu += wallcycle_stop(wcycle, ewcWAIT_GPU_NB_NL);
1262             }
1263             else
1264             {
1265                 wallcycle_start_nocount(wcycle, ewcFORCE);
1266                 do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::NonLocal, enbvClearFYes,
1267                              step, nrnb, wcycle);
1268                 wallcycle_stop(wcycle, ewcFORCE);
1269             }
1270
1271             nbv->atomdata_add_nbat_f_to_f(Nbnxm::AtomLocality::NonLocal,
1272                                           forceOut.f, wcycle);
1273         }
1274     }
1275
1276     if (havePPDomainDecomposition(cr))
1277     {
1278         /* We are done with the CPU compute.
1279          * We will now communicate the non-local forces.
1280          * If we use a GPU this will overlap with GPU work, so in that case
1281          * we do not close the DD force balancing region here.
1282          */
1283         ddBalanceRegionHandler.closeAfterForceComputationCpu();
1284
1285         if (bDoForces)
1286         {
1287             dd_move_f(cr->dd, force.unpaddedArrayRef(), fr->fshift, wcycle);
1288         }
1289     }
1290
1291     // With both nonbonded and PME offloaded a GPU on the same rank, we use
1292     // an alternating wait/reduction scheme.
1293     bool alternateGpuWait = (!c_disableAlternatingWait && useGpuPme && bUseGPU && !DOMAINDECOMP(cr));
1294     if (alternateGpuWait)
1295     {
1296         alternatePmeNbGpuWaitReduce(fr->nbv.get(), fr->pmedata, &force, &forceOut.forceWithVirial, fr->fshift, enerd,
1297                                     flags, pmeFlags, ppForceWorkload->haveGpuBondedWork, wcycle);
1298     }
1299
1300     if (!alternateGpuWait && useGpuPme)
1301     {
1302         pme_gpu_wait_and_reduce(fr->pmedata, pmeFlags, wcycle, &forceOut.forceWithVirial, enerd);
1303     }
1304
1305     /* Wait for local GPU NB outputs on the non-alternating wait path */
1306     if (!alternateGpuWait && bUseGPU)
1307     {
1308         /* Measured overhead on CUDA and OpenCL with(out) GPU sharing
1309          * is between 0.5 and 1.5 Mcycles. So 2 MCycles is an overestimate,
1310          * but even with a step of 0.1 ms the difference is less than 1%
1311          * of the step time.
1312          */
1313         const float gpuWaitApiOverheadMargin = 2e6f; /* cycles */
1314
1315         wallcycle_start(wcycle, ewcWAIT_GPU_NB_L);
1316         Nbnxm::gpu_wait_finish_task(nbv->gpu_nbv,
1317                                     flags, Nbnxm::AtomLocality::Local, ppForceWorkload->haveGpuBondedWork,
1318                                     enerd->grpp.ener[egLJSR], enerd->grpp.ener[egCOULSR],
1319                                     fr->fshift);
1320         float cycles_tmp = wallcycle_stop(wcycle, ewcWAIT_GPU_NB_L);
1321
1322         if (ddBalanceRegionHandler.useBalancingRegion())
1323         {
1324             DdBalanceRegionWaitedForGpu waitedForGpu = DdBalanceRegionWaitedForGpu::yes;
1325             if (bDoForces && cycles_tmp <= gpuWaitApiOverheadMargin)
1326             {
1327                 /* We measured few cycles, it could be that the kernel
1328                  * and transfer finished earlier and there was no actual
1329                  * wait time, only API call overhead.
1330                  * Then the actual time could be anywhere between 0 and
1331                  * cycles_wait_est. We will use half of cycles_wait_est.
1332                  */
1333                 waitedForGpu = DdBalanceRegionWaitedForGpu::no;
1334             }
1335             ddBalanceRegionHandler.closeAfterForceComputationGpu(cycles_wait_gpu, waitedForGpu);
1336         }
1337     }
1338
1339     if (fr->nbv->emulateGpu())
1340     {
1341         // NOTE: emulation kernel is not included in the balancing region,
1342         // but emulation mode does not target performance anyway
1343         wallcycle_start_nocount(wcycle, ewcFORCE);
1344         do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::Local,
1345                      DOMAINDECOMP(cr) ? enbvClearFNo : enbvClearFYes,
1346                      step, nrnb, wcycle);
1347         wallcycle_stop(wcycle, ewcFORCE);
1348     }
1349
1350     if (useGpuPme)
1351     {
1352         pme_gpu_reinit_computation(fr->pmedata, wcycle);
1353     }
1354
1355     if (bUseGPU)
1356     {
1357         /* now clear the GPU outputs while we finish the step on the CPU */
1358         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1359         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1360         Nbnxm::gpu_clear_outputs(nbv->gpu_nbv, flags);
1361
1362         if (nbv->pairlistSets().isDynamicPruningStepGpu(step))
1363         {
1364             nbv->dispatchPruneKernelGpu(step);
1365         }
1366         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1367         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1368     }
1369
1370     if (ppForceWorkload->haveGpuBondedWork && (flags & GMX_FORCE_ENERGY))
1371     {
1372         wallcycle_start(wcycle, ewcWAIT_GPU_BONDED);
1373         // in principle this should be included in the DD balancing region,
1374         // but generally it is infrequent so we'll omit it for the sake of
1375         // simpler code
1376         fr->gpuBonded->accumulateEnergyTerms(enerd);
1377         wallcycle_stop(wcycle, ewcWAIT_GPU_BONDED);
1378
1379         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1380         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_BONDED);
1381         fr->gpuBonded->clearEnergies();
1382         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1383         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1384     }
1385
1386     /* Do the nonbonded GPU (or emulation) force buffer reduction
1387      * on the non-alternating path. */
1388     if (bUseOrEmulGPU && !alternateGpuWait)
1389     {
1390         nbv->atomdata_add_nbat_f_to_f(Nbnxm::AtomLocality::Local,
1391                                       forceOut.f, wcycle);
1392     }
1393     if (DOMAINDECOMP(cr))
1394     {
1395         dd_force_flop_stop(cr->dd, nrnb);
1396     }
1397
1398     if (bDoForces)
1399     {
1400         /* If we have NoVirSum forces, but we do not calculate the virial,
1401          * we sum fr->f_novirsum=forceOut.f later.
1402          */
1403         if (vsite && !(fr->haveDirectVirialContributions && !(flags & GMX_FORCE_VIRIAL)))
1404         {
1405             spread_vsite_f(vsite, as_rvec_array(x.unpaddedArrayRef().data()), forceOut.f, fr->fshift, FALSE, nullptr, nrnb,
1406                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr, wcycle);
1407         }
1408
1409         if (flags & GMX_FORCE_VIRIAL)
1410         {
1411             /* Calculation of the virial must be done after vsites! */
1412             calc_virial(0, mdatoms->homenr, as_rvec_array(x.unpaddedArrayRef().data()), forceOut.f,
1413                         vir_force, graph, box, nrnb, fr, inputrec->ePBC);
1414         }
1415     }
1416
1417     if (PAR(cr) && !thisRankHasDuty(cr, DUTY_PME))
1418     {
1419         /* In case of node-splitting, the PP nodes receive the long-range
1420          * forces, virial and energy from the PME nodes here.
1421          */
1422         pme_receive_force_ener(cr, &forceOut.forceWithVirial, enerd, wcycle);
1423     }
1424
1425     if (bDoForces)
1426     {
1427         post_process_forces(cr, step, nrnb, wcycle,
1428                             top, box, as_rvec_array(x.unpaddedArrayRef().data()), forceOut.f, &forceOut.forceWithVirial,
1429                             vir_force, mdatoms, graph, fr, vsite,
1430                             flags);
1431     }
1432
1433     if (flags & GMX_FORCE_ENERGY)
1434     {
1435         /* Sum the potential energy terms from group contributions */
1436         sum_epot(&(enerd->grpp), enerd->term);
1437
1438         if (!EI_TPI(inputrec->eI))
1439         {
1440             checkPotentialEnergyValidity(step, *enerd, *inputrec);
1441         }
1442     }
1443 }
1444
1445 static void do_force_cutsGROUP(FILE *fplog,
1446                                const t_commrec *cr,
1447                                const gmx_multisim_t *ms,
1448                                const t_inputrec *inputrec,
1449                                gmx::Awh *awh,
1450                                gmx_enfrot *enforcedRotation,
1451                                int64_t step,
1452                                t_nrnb *nrnb,
1453                                gmx_wallcycle_t wcycle,
1454                                gmx_localtop_t *top,
1455                                const gmx_groups_t *groups,
1456                                matrix box, gmx::ArrayRefWithPadding<gmx::RVec> x,
1457                                history_t *hist,
1458                                gmx::ArrayRefWithPadding<gmx::RVec> force,
1459                                tensor vir_force,
1460                                const t_mdatoms *mdatoms,
1461                                gmx_enerdata_t *enerd,
1462                                t_fcdata *fcd,
1463                                real *lambda,
1464                                t_graph *graph,
1465                                t_forcerec *fr,
1466                                const gmx_vsite_t *vsite,
1467                                rvec mu_tot,
1468                                double t,
1469                                gmx_edsam *ed,
1470                                int flags,
1471                                const DDBalanceRegionHandler &ddBalanceRegionHandler)
1472 {
1473     int        cg0, cg1, i, j;
1474     double     mu[2*DIM];
1475     gmx_bool   bStateChanged, bNS, bFillGrid, bCalcCGCM;
1476     gmx_bool   bDoForces;
1477     float      cycles_pme;
1478
1479     const int  start  = 0;
1480     const int  homenr = mdatoms->homenr;
1481
1482     clear_mat(vir_force);
1483
1484     cg0 = 0;
1485     if (DOMAINDECOMP(cr))
1486     {
1487         cg1 = cr->dd->globalAtomGroupIndices.size();
1488     }
1489     else
1490     {
1491         cg1 = top->cgs.nr;
1492     }
1493     if (fr->n_tpi > 0)
1494     {
1495         cg1--;
1496     }
1497
1498     bStateChanged  = ((flags & GMX_FORCE_STATECHANGED) != 0);
1499     bNS            = ((flags & GMX_FORCE_NS) != 0);
1500     /* Should we perform the long-range nonbonded evaluation inside the neighborsearching? */
1501     bFillGrid      = (bNS && bStateChanged);
1502     bCalcCGCM      = (bFillGrid && !DOMAINDECOMP(cr));
1503     bDoForces      = ((flags & GMX_FORCE_FORCES) != 0);
1504
1505     if (bStateChanged)
1506     {
1507         update_forcerec(fr, box);
1508
1509         if (inputrecNeedMutot(inputrec))
1510         {
1511             /* Calculate total (local) dipole moment in a temporary common array.
1512              * This makes it possible to sum them over nodes faster.
1513              */
1514             calc_mu(start, homenr,
1515                     x.unpaddedArrayRef(), mdatoms->chargeA, mdatoms->chargeB, mdatoms->nChargePerturbed,
1516                     mu, mu+DIM);
1517         }
1518     }
1519
1520     if (fr->ePBC != epbcNONE)
1521     {
1522         /* Compute shift vectors every step,
1523          * because of pressure coupling or box deformation!
1524          */
1525         if ((flags & GMX_FORCE_DYNAMICBOX) && bStateChanged)
1526         {
1527             calc_shifts(box, fr->shift_vec);
1528         }
1529
1530         if (bCalcCGCM)
1531         {
1532             put_charge_groups_in_box(fplog, cg0, cg1, fr->ePBC, box,
1533                                      &(top->cgs), as_rvec_array(x.unpaddedArrayRef().data()), fr->cg_cm);
1534             inc_nrnb(nrnb, eNR_CGCM, homenr);
1535             inc_nrnb(nrnb, eNR_RESETX, cg1-cg0);
1536         }
1537         else if (EI_ENERGY_MINIMIZATION(inputrec->eI) && graph)
1538         {
1539             unshift_self(graph, box, as_rvec_array(x.unpaddedArrayRef().data()));
1540         }
1541     }
1542     else if (bCalcCGCM)
1543     {
1544         calc_cgcm(fplog, cg0, cg1, &(top->cgs), as_rvec_array(x.unpaddedArrayRef().data()), fr->cg_cm);
1545         inc_nrnb(nrnb, eNR_CGCM, homenr);
1546     }
1547
1548     if (bCalcCGCM && gmx_debug_at)
1549     {
1550         pr_rvecs(debug, 0, "cgcm", fr->cg_cm, top->cgs.nr);
1551     }
1552
1553 #if GMX_MPI
1554     if (!thisRankHasDuty(cr, DUTY_PME))
1555     {
1556         /* Send particle coordinates to the pme nodes.
1557          * Since this is only implemented for domain decomposition
1558          * and domain decomposition does not use the graph,
1559          * we do not need to worry about shifting.
1560          */
1561         gmx_pme_send_coordinates(cr, box, as_rvec_array(x.unpaddedArrayRef().data()),
1562                                  lambda[efptCOUL], lambda[efptVDW],
1563                                  (flags & (GMX_FORCE_VIRIAL | GMX_FORCE_ENERGY)) != 0,
1564                                  step, wcycle);
1565     }
1566 #endif /* GMX_MPI */
1567
1568     /* Communicate coordinates and sum dipole if necessary */
1569     if (DOMAINDECOMP(cr))
1570     {
1571         dd_move_x(cr->dd, box, x.unpaddedArrayRef(), wcycle);
1572
1573         /* No GPU support, no move_x overlap, so reopen the balance region here */
1574         ddBalanceRegionHandler.reopenRegionCpu();
1575     }
1576
1577     if (inputrecNeedMutot(inputrec))
1578     {
1579         if (bStateChanged)
1580         {
1581             if (PAR(cr))
1582             {
1583                 gmx_sumd(2*DIM, mu, cr);
1584
1585                 ddBalanceRegionHandler.reopenRegionCpu();
1586             }
1587             for (i = 0; i < 2; i++)
1588             {
1589                 for (j = 0; j < DIM; j++)
1590                 {
1591                     fr->mu_tot[i][j] = mu[i*DIM + j];
1592                 }
1593             }
1594         }
1595         if (fr->efep == efepNO)
1596         {
1597             copy_rvec(fr->mu_tot[0], mu_tot);
1598         }
1599         else
1600         {
1601             for (j = 0; j < DIM; j++)
1602             {
1603                 mu_tot[j] =
1604                     (1.0 - lambda[efptCOUL])*fr->mu_tot[0][j] + lambda[efptCOUL]*fr->mu_tot[1][j];
1605             }
1606         }
1607     }
1608
1609     /* Reset energies */
1610     reset_enerdata(enerd);
1611     clear_rvecs(SHIFTS, fr->fshift);
1612
1613     if (bNS)
1614     {
1615         wallcycle_start(wcycle, ewcNS);
1616
1617         if (graph && bStateChanged)
1618         {
1619             /* Calculate intramolecular shift vectors to make molecules whole */
1620             mk_mshift(fplog, graph, fr->ePBC, box, as_rvec_array(x.unpaddedArrayRef().data()));
1621         }
1622
1623         /* Do the actual neighbour searching */
1624         ns(fplog, fr, box,
1625            groups, top, mdatoms,
1626            cr, nrnb, bFillGrid);
1627
1628         wallcycle_stop(wcycle, ewcNS);
1629     }
1630
1631     if (DOMAINDECOMP(cr) && !thisRankHasDuty(cr, DUTY_PME))
1632     {
1633         wallcycle_start(wcycle, ewcPPDURINGPME);
1634         dd_force_flop_start(cr->dd, nrnb);
1635     }
1636
1637     if (inputrec->bRot)
1638     {
1639         wallcycle_start(wcycle, ewcROT);
1640         do_rotation(cr, enforcedRotation, box, as_rvec_array(x.unpaddedArrayRef().data()), t, step, bNS);
1641         wallcycle_stop(wcycle, ewcROT);
1642     }
1643
1644
1645     /* Start the force cycle counter.
1646      * Note that a different counter is used for dynamic load balancing.
1647      */
1648     wallcycle_start(wcycle, ewcFORCE);
1649
1650     // set up and clear force outputs
1651     struct ForceOutputs forceOut = setupForceOutputs(fr, *inputrec, force, bDoForces, ((flags & GMX_FORCE_VIRIAL) != 0), wcycle);
1652
1653     if (inputrec->bPull && pull_have_constraint(inputrec->pull_work))
1654     {
1655         clear_pull_forces(inputrec->pull_work);
1656     }
1657
1658     /* update QMMMrec, if necessary */
1659     if (fr->bQMMM)
1660     {
1661         update_QMMMrec(cr, fr, as_rvec_array(x.unpaddedArrayRef().data()), mdatoms, box);
1662     }
1663
1664     /* Compute the bonded and non-bonded energies and optionally forces */
1665     do_force_lowlevel(fr, inputrec, &(top->idef),
1666                       cr, ms, nrnb, wcycle, mdatoms,
1667                       as_rvec_array(x.unpaddedArrayRef().data()), hist, forceOut.f, &forceOut.forceWithVirial, enerd, fcd,
1668                       box, inputrec->fepvals, lambda,
1669                       graph, &(top->excls), fr->mu_tot,
1670                       flags,
1671                       &cycles_pme, ddBalanceRegionHandler);
1672
1673     wallcycle_stop(wcycle, ewcFORCE);
1674
1675     if (DOMAINDECOMP(cr))
1676     {
1677         dd_force_flop_stop(cr->dd, nrnb);
1678
1679         ddBalanceRegionHandler.closeAfterForceComputationCpu();
1680     }
1681
1682     computeSpecialForces(fplog, cr, inputrec, awh, enforcedRotation,
1683                          step, t, wcycle,
1684                          fr->forceProviders, box, x.unpaddedArrayRef(), mdatoms, lambda,
1685                          flags, &forceOut.forceWithVirial, enerd,
1686                          ed, bNS);
1687
1688     if (bDoForces)
1689     {
1690         /* Communicate the forces */
1691         if (DOMAINDECOMP(cr))
1692         {
1693             dd_move_f(cr->dd, force.unpaddedArrayRef(), fr->fshift, wcycle);
1694             /* Do we need to communicate the separate force array
1695              * for terms that do not contribute to the single sum virial?
1696              * Position restraints and electric fields do not introduce
1697              * inter-cg forces, only full electrostatics methods do.
1698              * When we do not calculate the virial, fr->f_novirsum = forceOut.f,
1699              * so we have already communicated these forces.
1700              */
1701             if (EEL_FULL(fr->ic->eeltype) && cr->dd->n_intercg_excl &&
1702                 (flags & GMX_FORCE_VIRIAL))
1703             {
1704                 dd_move_f(cr->dd, forceOut.forceWithVirial.force_, nullptr, wcycle);
1705             }
1706         }
1707
1708         /* If we have NoVirSum forces, but we do not calculate the virial,
1709          * we sum fr->f_novirsum=forceOut.f later.
1710          */
1711         if (vsite && !(fr->haveDirectVirialContributions && !(flags & GMX_FORCE_VIRIAL)))
1712         {
1713             spread_vsite_f(vsite, as_rvec_array(x.unpaddedArrayRef().data()), forceOut.f, fr->fshift, FALSE, nullptr, nrnb,
1714                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr, wcycle);
1715         }
1716
1717         if (flags & GMX_FORCE_VIRIAL)
1718         {
1719             /* Calculation of the virial must be done after vsites! */
1720             calc_virial(0, mdatoms->homenr, as_rvec_array(x.unpaddedArrayRef().data()), forceOut.f,
1721                         vir_force, graph, box, nrnb, fr, inputrec->ePBC);
1722         }
1723     }
1724
1725     if (PAR(cr) && !thisRankHasDuty(cr, DUTY_PME))
1726     {
1727         /* In case of node-splitting, the PP nodes receive the long-range
1728          * forces, virial and energy from the PME nodes here.
1729          */
1730         pme_receive_force_ener(cr, &forceOut.forceWithVirial, enerd, wcycle);
1731     }
1732
1733     if (bDoForces)
1734     {
1735         post_process_forces(cr, step, nrnb, wcycle,
1736                             top, box, as_rvec_array(x.unpaddedArrayRef().data()), forceOut.f, &forceOut.forceWithVirial,
1737                             vir_force, mdatoms, graph, fr, vsite,
1738                             flags);
1739     }
1740
1741     if (flags & GMX_FORCE_ENERGY)
1742     {
1743         /* Sum the potential energy terms from group contributions */
1744         sum_epot(&(enerd->grpp), enerd->term);
1745
1746         if (!EI_TPI(inputrec->eI))
1747         {
1748             checkPotentialEnergyValidity(step, *enerd, *inputrec);
1749         }
1750     }
1751
1752 }
1753
1754 void do_force(FILE                                     *fplog,
1755               const t_commrec                          *cr,
1756               const gmx_multisim_t                     *ms,
1757               const t_inputrec                         *inputrec,
1758               gmx::Awh                                 *awh,
1759               gmx_enfrot                               *enforcedRotation,
1760               int64_t                                   step,
1761               t_nrnb                                   *nrnb,
1762               gmx_wallcycle_t                           wcycle,
1763               gmx_localtop_t                           *top,
1764               const gmx_groups_t                       *groups,
1765               matrix                                    box,
1766               gmx::ArrayRefWithPadding<gmx::RVec>       x,     //NOLINT(performance-unnecessary-value-param)
1767               history_t                                *hist,
1768               gmx::ArrayRefWithPadding<gmx::RVec>       force, //NOLINT(performance-unnecessary-value-param)
1769               tensor                                    vir_force,
1770               const t_mdatoms                          *mdatoms,
1771               gmx_enerdata_t                           *enerd,
1772               t_fcdata                                 *fcd,
1773               gmx::ArrayRef<real>                       lambda,
1774               t_graph                                  *graph,
1775               t_forcerec                               *fr,
1776               gmx::PpForceWorkload                     *ppForceWorkload,
1777               const gmx_vsite_t                        *vsite,
1778               rvec                                      mu_tot,
1779               double                                    t,
1780               gmx_edsam                                *ed,
1781               int                                       flags,
1782               const DDBalanceRegionHandler             &ddBalanceRegionHandler)
1783 {
1784     /* modify force flag if not doing nonbonded */
1785     if (!fr->bNonbonded)
1786     {
1787         flags &= ~GMX_FORCE_NONBONDED;
1788     }
1789
1790     switch (inputrec->cutoff_scheme)
1791     {
1792         case ecutsVERLET:
1793             do_force_cutsVERLET(fplog, cr, ms, inputrec,
1794                                 awh, enforcedRotation, step, nrnb, wcycle,
1795                                 top,
1796                                 groups,
1797                                 box, x, hist,
1798                                 force, vir_force,
1799                                 mdatoms,
1800                                 enerd, fcd,
1801                                 lambda.data(), graph,
1802                                 fr,
1803                                 ppForceWorkload,
1804                                 fr->ic,
1805                                 vsite, mu_tot,
1806                                 t, ed,
1807                                 flags,
1808                                 ddBalanceRegionHandler);
1809             break;
1810         case ecutsGROUP:
1811             do_force_cutsGROUP(fplog, cr, ms, inputrec,
1812                                awh, enforcedRotation, step, nrnb, wcycle,
1813                                top,
1814                                groups,
1815                                box, x, hist,
1816                                force, vir_force,
1817                                mdatoms,
1818                                enerd, fcd,
1819                                lambda.data(), graph,
1820                                fr, vsite, mu_tot,
1821                                t, ed,
1822                                flags,
1823                                ddBalanceRegionHandler);
1824             break;
1825         default:
1826             gmx_incons("Invalid cut-off scheme passed!");
1827     }
1828
1829     /* In case we don't have constraints and are using GPUs, the next balancing
1830      * region starts here.
1831      * Some "special" work at the end of do_force_cuts?, such as vsite spread,
1832      * virial calculation and COM pulling, is not thus not included in
1833      * the balance timing, which is ok as most tasks do communication.
1834      */
1835     ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::no);
1836 }
1837
1838
1839 void do_constrain_first(FILE *fplog, gmx::Constraints *constr,
1840                         const t_inputrec *ir, const t_mdatoms *md,
1841                         t_state *state)
1842 {
1843     int             i, m, start, end;
1844     int64_t         step;
1845     real            dt = ir->delta_t;
1846     real            dvdl_dum;
1847     rvec           *savex;
1848
1849     /* We need to allocate one element extra, since we might use
1850      * (unaligned) 4-wide SIMD loads to access rvec entries.
1851      */
1852     snew(savex, state->natoms + 1);
1853
1854     start = 0;
1855     end   = md->homenr;
1856
1857     if (debug)
1858     {
1859         fprintf(debug, "vcm: start=%d, homenr=%d, end=%d\n",
1860                 start, md->homenr, end);
1861     }
1862     /* Do a first constrain to reset particles... */
1863     step = ir->init_step;
1864     if (fplog)
1865     {
1866         char buf[STEPSTRSIZE];
1867         fprintf(fplog, "\nConstraining the starting coordinates (step %s)\n",
1868                 gmx_step_str(step, buf));
1869     }
1870     dvdl_dum = 0;
1871
1872     /* constrain the current position */
1873     constr->apply(TRUE, FALSE,
1874                   step, 0, 1.0,
1875                   state->x.rvec_array(), state->x.rvec_array(), nullptr,
1876                   state->box,
1877                   state->lambda[efptBONDED], &dvdl_dum,
1878                   nullptr, nullptr, gmx::ConstraintVariable::Positions);
1879     if (EI_VV(ir->eI))
1880     {
1881         /* constrain the inital velocity, and save it */
1882         /* also may be useful if we need the ekin from the halfstep for velocity verlet */
1883         constr->apply(TRUE, FALSE,
1884                       step, 0, 1.0,
1885                       state->x.rvec_array(), state->v.rvec_array(), state->v.rvec_array(),
1886                       state->box,
1887                       state->lambda[efptBONDED], &dvdl_dum,
1888                       nullptr, nullptr, gmx::ConstraintVariable::Velocities);
1889     }
1890     /* constrain the inital velocities at t-dt/2 */
1891     if (EI_STATE_VELOCITY(ir->eI) && ir->eI != eiVV)
1892     {
1893         auto x = makeArrayRef(state->x).subArray(start, end);
1894         auto v = makeArrayRef(state->v).subArray(start, end);
1895         for (i = start; (i < end); i++)
1896         {
1897             for (m = 0; (m < DIM); m++)
1898             {
1899                 /* Reverse the velocity */
1900                 v[i][m] = -v[i][m];
1901                 /* Store the position at t-dt in buf */
1902                 savex[i][m] = x[i][m] + dt*v[i][m];
1903             }
1904         }
1905         /* Shake the positions at t=-dt with the positions at t=0
1906          * as reference coordinates.
1907          */
1908         if (fplog)
1909         {
1910             char buf[STEPSTRSIZE];
1911             fprintf(fplog, "\nConstraining the coordinates at t0-dt (step %s)\n",
1912                     gmx_step_str(step, buf));
1913         }
1914         dvdl_dum = 0;
1915         constr->apply(TRUE, FALSE,
1916                       step, -1, 1.0,
1917                       state->x.rvec_array(), savex, nullptr,
1918                       state->box,
1919                       state->lambda[efptBONDED], &dvdl_dum,
1920                       state->v.rvec_array(), nullptr, gmx::ConstraintVariable::Positions);
1921
1922         for (i = start; i < end; i++)
1923         {
1924             for (m = 0; m < DIM; m++)
1925             {
1926                 /* Re-reverse the velocities */
1927                 v[i][m] = -v[i][m];
1928             }
1929         }
1930     }
1931     sfree(savex);
1932 }
1933
1934 void put_atoms_in_box_omp(int ePBC, const matrix box, gmx::ArrayRef<gmx::RVec> x)
1935 {
1936     int t, nth;
1937     nth = gmx_omp_nthreads_get(emntDefault);
1938
1939 #pragma omp parallel for num_threads(nth) schedule(static)
1940     for (t = 0; t < nth; t++)
1941     {
1942         try
1943         {
1944             size_t natoms = x.size();
1945             size_t offset = (natoms*t    )/nth;
1946             size_t len    = (natoms*(t + 1))/nth - offset;
1947             put_atoms_in_box(ePBC, box, x.subArray(offset, len));
1948         }
1949         GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1950     }
1951 }
1952
1953 void initialize_lambdas(FILE               *fplog,
1954                         const t_inputrec   &ir,
1955                         bool                isMaster,
1956                         int                *fep_state,
1957                         gmx::ArrayRef<real> lambda,
1958                         double             *lam0)
1959 {
1960     /* TODO: Clean up initialization of fep_state and lambda in
1961        t_state.  This function works, but could probably use a logic
1962        rewrite to keep all the different types of efep straight. */
1963
1964     if ((ir.efep == efepNO) && (!ir.bSimTemp))
1965     {
1966         return;
1967     }
1968
1969     const t_lambda *fep = ir.fepvals;
1970     if (isMaster)
1971     {
1972         *fep_state = fep->init_fep_state; /* this might overwrite the checkpoint
1973                                              if checkpoint is set -- a kludge is in for now
1974                                              to prevent this.*/
1975     }
1976
1977     for (int i = 0; i < efptNR; i++)
1978     {
1979         double thisLambda;
1980         /* overwrite lambda state with init_lambda for now for backwards compatibility */
1981         if (fep->init_lambda >= 0) /* if it's -1, it was never initialized */
1982         {
1983             thisLambda = fep->init_lambda;
1984         }
1985         else
1986         {
1987             thisLambda = fep->all_lambda[i][fep->init_fep_state];
1988         }
1989         if (isMaster)
1990         {
1991             lambda[i] = thisLambda;
1992         }
1993         if (lam0 != nullptr)
1994         {
1995             lam0[i] = thisLambda;
1996         }
1997     }
1998     if (ir.bSimTemp)
1999     {
2000         /* need to rescale control temperatures to match current state */
2001         for (int i = 0; i < ir.opts.ngtc; i++)
2002         {
2003             if (ir.opts.ref_t[i] > 0)
2004             {
2005                 ir.opts.ref_t[i] = ir.simtempvals->temperatures[fep->init_fep_state];
2006             }
2007         }
2008     }
2009
2010     /* Send to the log the information on the current lambdas */
2011     if (fplog != nullptr)
2012     {
2013         fprintf(fplog, "Initial vector of lambda components:[ ");
2014         for (const auto &l : lambda)
2015         {
2016             fprintf(fplog, "%10.4f ", l);
2017         }
2018         fprintf(fplog, "]\n");
2019     }
2020 }