src/gromacs/mdlib/sim_util.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
   5  * Copyright (c) 2001-2004, The GROMACS development team.
   6  * Copyright (c) 2013,2014,2015,2016,2017,2018,2019, by the GROMACS development team, led by
   7  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   8  * and including many others, as listed in the AUTHORS file in the
   9  * top-level source directory and at http://www.gromacs.org.
  10  *
  11  * GROMACS is free software; you can redistribute it and/or
  12  * modify it under the terms of the GNU Lesser General Public License
  13  * as published by the Free Software Foundation; either version 2.1
  14  * of the License, or (at your option) any later version.
  15  *
  16  * GROMACS is distributed in the hope that it will be useful,
  17  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  18  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  19  * Lesser General Public License for more details.
  20  *
  21  * You should have received a copy of the GNU Lesser General Public
  22  * License along with GROMACS; if not, see
  23  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  24  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  25  *
  26  * If you want to redistribute modifications to GROMACS, please
  27  * consider that scientific software is very special. Version
  28  * control is crucial - bugs must be traceable. We will be happy to
  29  * consider code for inclusion in the official distribution, but
  30  * derived work must not be called official GROMACS. Details are found
  31  * in the README & COPYING files - if they are missing, get the
  32  * official version at http://www.gromacs.org.
  33  *
  34  * To help us fund GROMACS development, we humbly ask that you cite
  35  * the research papers on the package. Check out http://www.gromacs.org.
  36  */
  37 #include "gmxpre.h"
  38
  39 #include "config.h"
  40
  41 #include <cmath>
  42 #include <cstdint>
  43 #include <cstdio>
  44 #include <cstring>
  45
  46 #include <array>
  47
  48 #include "gromacs/awh/awh.h"
  49 #include "gromacs/domdec/dlbtiming.h"
  50 #include "gromacs/domdec/domdec.h"
  51 #include "gromacs/domdec/domdec_struct.h"
  52 #include "gromacs/domdec/partition.h"
  53 #include "gromacs/essentialdynamics/edsam.h"
  54 #include "gromacs/ewald/pme.h"
  55 #include "gromacs/gmxlib/chargegroup.h"
  56 #include "gromacs/gmxlib/network.h"
  57 #include "gromacs/gmxlib/nrnb.h"
  58 #include "gromacs/gmxlib/nonbonded/nb_free_energy.h"
  59 #include "gromacs/gmxlib/nonbonded/nb_kernel.h"
  60 #include "gromacs/gmxlib/nonbonded/nonbonded.h"
  61 #include "gromacs/gpu_utils/gpu_utils.h"
  62 #include "gromacs/imd/imd.h"
  63 #include "gromacs/listed_forces/bonded.h"
  64 #include "gromacs/listed_forces/disre.h"
  65 #include "gromacs/listed_forces/gpubonded.h"
  66 #include "gromacs/listed_forces/listed_forces.h"
  67 #include "gromacs/listed_forces/manage_threading.h"
  68 #include "gromacs/listed_forces/orires.h"
  69 #include "gromacs/math/arrayrefwithpadding.h"
  70 #include "gromacs/math/functions.h"
  71 #include "gromacs/math/units.h"
  72 #include "gromacs/math/vec.h"
  73 #include "gromacs/math/vecdump.h"
  74 #include "gromacs/mdlib/calcmu.h"
  75 #include "gromacs/mdlib/calcvir.h"
  76 #include "gromacs/mdlib/constr.h"
  77 #include "gromacs/mdlib/enerdata_utils.h"
  78 #include "gromacs/mdlib/force.h"
  79 #include "gromacs/mdlib/forcerec.h"
  80 #include "gromacs/mdlib/gmx_omp_nthreads.h"
  81 #include "gromacs/mdlib/ppforceworkload.h"
  82 #include "gromacs/mdlib/qmmm.h"
  83 #include "gromacs/mdlib/update.h"
  84 #include "gromacs/mdtypes/commrec.h"
  85 #include "gromacs/mdtypes/enerdata.h"
  86 #include "gromacs/mdtypes/forceoutput.h"
  87 #include "gromacs/mdtypes/iforceprovider.h"
  88 #include "gromacs/mdtypes/inputrec.h"
  89 #include "gromacs/mdtypes/md_enums.h"
  90 #include "gromacs/mdtypes/state.h"
  91 #include "gromacs/nbnxm/atomdata.h"
  92 #include "gromacs/nbnxm/gpu_data_mgmt.h"
  93 #include "gromacs/nbnxm/nbnxm.h"
  94 #include "gromacs/pbcutil/ishift.h"
  95 #include "gromacs/pbcutil/mshift.h"
  96 #include "gromacs/pbcutil/pbc.h"
  97 #include "gromacs/pulling/pull.h"
  98 #include "gromacs/pulling/pull_rotation.h"
  99 #include "gromacs/timing/cyclecounter.h"
 100 #include "gromacs/timing/gpu_timing.h"
 101 #include "gromacs/timing/wallcycle.h"
 102 #include "gromacs/timing/wallcyclereporting.h"
 103 #include "gromacs/timing/walltime_accounting.h"
 104 #include "gromacs/topology/topology.h"
 105 #include "gromacs/utility/arrayref.h"
 106 #include "gromacs/utility/basedefinitions.h"
 107 #include "gromacs/utility/cstringutil.h"
 108 #include "gromacs/utility/exceptions.h"
 109 #include "gromacs/utility/fatalerror.h"
 110 #include "gromacs/utility/gmxassert.h"
 111 #include "gromacs/utility/gmxmpi.h"
 112 #include "gromacs/utility/logger.h"
 113 #include "gromacs/utility/smalloc.h"
 114 #include "gromacs/utility/strconvert.h"
 115 #include "gromacs/utility/sysinfo.h"
 116
 117 // TODO: this environment variable allows us to verify before release
 118 // that on less common architectures the total cost of polling is not larger than
 119 // a blocking wait (so polling does not introduce overhead when the static
 120 // PME-first ordering would suffice).
 121 static const bool c_disableAlternatingWait = (getenv("GMX_DISABLE_ALTERNATING_GPU_WAIT") != nullptr);
 122
 123 // environment variable to enable GPU buffer ops, to alow incremental and optional
 124 // introduction of this functionality.
 125 // TODO eventially tie this in with other existing GPU flags.
 126 static const bool c_enableGpuBufOps = (getenv("GMX_USE_GPU_BUFFER_OPS") != nullptr);
 127
 128
 129
 130 static void sum_forces(rvec f[], gmx::ArrayRef<const gmx::RVec> forceToAdd)
 131 {
 132     const int      end = forceToAdd.size();
 133
 134     int gmx_unused nt = gmx_omp_nthreads_get(emntDefault);
 135 #pragma omp parallel for num_threads(nt) schedule(static)
 136     for (int i = 0; i < end; i++)
 137     {
 138         rvec_inc(f[i], forceToAdd[i]);
 139     }
 140 }
 141
 142 static void calc_virial(int start, int homenr, const rvec x[], const rvec f[],
 143                         tensor vir_part, const t_graph *graph, const matrix box,
 144                         t_nrnb *nrnb, const t_forcerec *fr, int ePBC)
 145 {
 146     /* The short-range virial from surrounding boxes */
 147     calc_vir(SHIFTS, fr->shift_vec, fr->fshift, vir_part, ePBC == epbcSCREW, box);
 148     inc_nrnb(nrnb, eNR_VIRIAL, SHIFTS);
 149
 150     /* Calculate partial virial, for local atoms only, based on short range.
 151      * Total virial is computed in global_stat, called from do_md
 152      */
 153     f_calc_vir(start, start+homenr, x, f, vir_part, graph, box);
 154     inc_nrnb(nrnb, eNR_VIRIAL, homenr);
 155
 156     if (debug)
 157     {
 158         pr_rvecs(debug, 0, "vir_part", vir_part, DIM);
 159     }
 160 }
 161
 162 static void pull_potential_wrapper(const t_commrec *cr,
 163                                    const t_inputrec *ir,
 164                                    const matrix box, gmx::ArrayRef<const gmx::RVec> x,
 165                                    gmx::ForceWithVirial *force,
 166                                    const t_mdatoms *mdatoms,
 167                                    gmx_enerdata_t *enerd,
 168                                    pull_t *pull_work,
 169                                    const real *lambda,
 170                                    double t,
 171                                    gmx_wallcycle_t wcycle)
 172 {
 173     t_pbc  pbc;
 174     real   dvdl;
 175
 176     /* Calculate the center of mass forces, this requires communication,
 177      * which is why pull_potential is called close to other communication.
 178      */
 179     wallcycle_start(wcycle, ewcPULLPOT);
 180     set_pbc(&pbc, ir->ePBC, box);
 181     dvdl                     = 0;
 182     enerd->term[F_COM_PULL] +=
 183         pull_potential(pull_work, mdatoms, &pbc,
 184                        cr, t, lambda[efptRESTRAINT], as_rvec_array(x.data()), force, &dvdl);
 185     enerd->dvdl_lin[efptRESTRAINT] += dvdl;
 186     wallcycle_stop(wcycle, ewcPULLPOT);
 187 }
 188
 189 static void pme_receive_force_ener(const t_commrec      *cr,
 190                                    gmx::ForceWithVirial *forceWithVirial,
 191                                    gmx_enerdata_t       *enerd,
 192                                    gmx_wallcycle_t       wcycle)
 193 {
 194     real   e_q, e_lj, dvdl_q, dvdl_lj;
 195     float  cycles_ppdpme, cycles_seppme;
 196
 197     cycles_ppdpme = wallcycle_stop(wcycle, ewcPPDURINGPME);
 198     dd_cycles_add(cr->dd, cycles_ppdpme, ddCyclPPduringPME);
 199
 200     /* In case of node-splitting, the PP nodes receive the long-range
 201      * forces, virial and energy from the PME nodes here.
 202      */
 203     wallcycle_start(wcycle, ewcPP_PMEWAITRECVF);
 204     dvdl_q  = 0;
 205     dvdl_lj = 0;
 206     gmx_pme_receive_f(cr, forceWithVirial, &e_q, &e_lj, &dvdl_q, &dvdl_lj,
 207                       &cycles_seppme);
 208     enerd->term[F_COUL_RECIP] += e_q;
 209     enerd->term[F_LJ_RECIP]   += e_lj;
 210     enerd->dvdl_lin[efptCOUL] += dvdl_q;
 211     enerd->dvdl_lin[efptVDW]  += dvdl_lj;
 212
 213     if (wcycle)
 214     {
 215         dd_cycles_add(cr->dd, cycles_seppme, ddCyclPME);
 216     }
 217     wallcycle_stop(wcycle, ewcPP_PMEWAITRECVF);
 218 }
 219
 220 static void print_large_forces(FILE            *fp,
 221                                const t_mdatoms *md,
 222                                const t_commrec *cr,
 223                                int64_t          step,
 224                                real             forceTolerance,
 225                                const rvec      *x,
 226                                const rvec      *f)
 227 {
 228     real           force2Tolerance = gmx::square(forceTolerance);
 229     gmx::index     numNonFinite    = 0;
 230     for (int i = 0; i < md->homenr; i++)
 231     {
 232         real force2    = norm2(f[i]);
 233         bool nonFinite = !std::isfinite(force2);
 234         if (force2 >= force2Tolerance || nonFinite)
 235         {
 236             fprintf(fp, "step %" PRId64 " atom %6d  x %8.3f %8.3f %8.3f  force %12.5e\n",
 237                     step,
 238                     ddglatnr(cr->dd, i), x[i][XX], x[i][YY], x[i][ZZ], std::sqrt(force2));
 239         }
 240         if (nonFinite)
 241         {
 242             numNonFinite++;
 243         }
 244     }
 245     if (numNonFinite > 0)
 246     {
 247         /* Note that with MPI this fatal call on one rank might interrupt
 248          * the printing on other ranks. But we can only avoid that with
 249          * an expensive MPI barrier that we would need at each step.
 250          */
 251         gmx_fatal(FARGS, "At step %" PRId64 " detected non-finite forces on %td atoms", step, numNonFinite);
 252     }
 253 }
 254
 255 static void post_process_forces(const t_commrec           *cr,
 256                                 int64_t                    step,
 257                                 t_nrnb                    *nrnb,
 258                                 gmx_wallcycle_t            wcycle,
 259                                 const gmx_localtop_t      *top,
 260                                 const matrix               box,
 261                                 const rvec                 x[],
 262                                 rvec                       f[],
 263                                 gmx::ForceWithVirial      *forceWithVirial,
 264                                 tensor                     vir_force,
 265                                 const t_mdatoms           *mdatoms,
 266                                 const t_graph             *graph,
 267                                 const t_forcerec          *fr,
 268                                 const gmx_vsite_t         *vsite,
 269                                 int                        flags)
 270 {
 271     if (fr->haveDirectVirialContributions)
 272     {
 273         rvec *fDirectVir = as_rvec_array(forceWithVirial->force_.data());
 274
 275         if (vsite)
 276         {
 277             /* Spread the mesh force on virtual sites to the other particles...
 278              * This is parallellized. MPI communication is performed
 279              * if the constructing atoms aren't local.
 280              */
 281             matrix virial = { { 0 } };
 282             spread_vsite_f(vsite, x, fDirectVir, nullptr,
 283                            (flags & GMX_FORCE_VIRIAL) != 0, virial,
 284                            nrnb,
 285                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr, wcycle);
 286             forceWithVirial->addVirialContribution(virial);
 287         }
 288
 289         if (flags & GMX_FORCE_VIRIAL)
 290         {
 291             /* Now add the forces, this is local */
 292             sum_forces(f, forceWithVirial->force_);
 293
 294             /* Add the direct virial contributions */
 295             GMX_ASSERT(forceWithVirial->computeVirial_, "forceWithVirial should request virial computation when we request the virial");
 296             m_add(vir_force, forceWithVirial->getVirial(), vir_force);
 297
 298             if (debug)
 299             {
 300                 pr_rvecs(debug, 0, "vir_force", vir_force, DIM);
 301             }
 302         }
 303     }
 304
 305     if (fr->print_force >= 0)
 306     {
 307         print_large_forces(stderr, mdatoms, cr, step, fr->print_force, x, f);
 308     }
 309 }
 310
 311 static void do_nb_verlet(t_forcerec                       *fr,
 312                          const interaction_const_t        *ic,
 313                          gmx_enerdata_t                   *enerd,
 314                          const int                         flags,
 315                          const Nbnxm::InteractionLocality  ilocality,
 316                          const int                         clearF,
 317                          const int64_t                     step,
 318                          t_nrnb                           *nrnb,
 319                          gmx_wallcycle_t                   wcycle)
 320 {
 321     if (!(flags & GMX_FORCE_NONBONDED))
 322     {
 323         /* skip non-bonded calculation */
 324         return;
 325     }
 326
 327     nonbonded_verlet_t *nbv  = fr->nbv.get();
 328
 329     /* GPU kernel launch overhead is already timed separately */
 330     if (fr->cutoff_scheme != ecutsVERLET)
 331     {
 332         gmx_incons("Invalid cut-off scheme passed!");
 333     }
 334
 335     if (!nbv->useGpu())
 336     {
 337         /* When dynamic pair-list  pruning is requested, we need to prune
 338          * at nstlistPrune steps.
 339          */
 340         if (nbv->isDynamicPruningStepCpu(step))
 341         {
 342             /* Prune the pair-list beyond fr->ic->rlistPrune using
 343              * the current coordinates of the atoms.
 344              */
 345             wallcycle_sub_start(wcycle, ewcsNONBONDED_PRUNING);
 346             nbv->dispatchPruneKernelCpu(ilocality, fr->shift_vec);
 347             wallcycle_sub_stop(wcycle, ewcsNONBONDED_PRUNING);
 348         }
 349
 350         wallcycle_sub_start(wcycle, ewcsNONBONDED);
 351     }
 352
 353     nbv->dispatchNonbondedKernel(ilocality, *ic, flags, clearF, *fr, enerd, nrnb, wcycle);
 354
 355     if (!nbv->useGpu())
 356     {
 357         wallcycle_sub_stop(wcycle, ewcsNONBONDED);
 358     }
 359 }
 360
 361 static inline void clear_rvecs_omp(int n, rvec v[])
 362 {
 363     int nth = gmx_omp_nthreads_get_simple_rvec_task(emntDefault, n);
 364
 365     /* Note that we would like to avoid this conditional by putting it
 366      * into the omp pragma instead, but then we still take the full
 367      * omp parallel for overhead (at least with gcc5).
 368      */
 369     if (nth == 1)
 370     {
 371         for (int i = 0; i < n; i++)
 372         {
 373             clear_rvec(v[i]);
 374         }
 375     }
 376     else
 377     {
 378 #pragma omp parallel for num_threads(nth) schedule(static)
 379         for (int i = 0; i < n; i++)
 380         {
 381             clear_rvec(v[i]);
 382         }
 383     }
 384 }
 385
 386 /*! \brief Return an estimate of the average kinetic energy or 0 when unreliable
 387  *
 388  * \param groupOptions  Group options, containing T-coupling options
 389  */
 390 static real averageKineticEnergyEstimate(const t_grpopts &groupOptions)
 391 {
 392     real nrdfCoupled   = 0;
 393     real nrdfUncoupled = 0;
 394     real kineticEnergy = 0;
 395     for (int g = 0; g < groupOptions.ngtc; g++)
 396     {
 397         if (groupOptions.tau_t[g] >= 0)
 398         {
 399             nrdfCoupled   += groupOptions.nrdf[g];
 400             kineticEnergy += groupOptions.nrdf[g]*0.5*groupOptions.ref_t[g]*BOLTZ;
 401         }
 402         else
 403         {
 404             nrdfUncoupled += groupOptions.nrdf[g];
 405         }
 406     }
 407
 408     /* This conditional with > also catches nrdf=0 */
 409     if (nrdfCoupled > nrdfUncoupled)
 410     {
 411         return kineticEnergy*(nrdfCoupled + nrdfUncoupled)/nrdfCoupled;
 412     }
 413     else
 414     {
 415         return 0;
 416     }
 417 }
 418
 419 /*! \brief This routine checks that the potential energy is finite.
 420  *
 421  * Always checks that the potential energy is finite. If step equals
 422  * inputrec.init_step also checks that the magnitude of the potential energy
 423  * is reasonable. Terminates with a fatal error when a check fails.
 424  * Note that passing this check does not guarantee finite forces,
 425  * since those use slightly different arithmetics. But in most cases
 426  * there is just a narrow coordinate range where forces are not finite
 427  * and energies are finite.
 428  *
 429  * \param[in] step      The step number, used for checking and printing
 430  * \param[in] enerd     The energy data; the non-bonded group energies need to be added to enerd.term[F_EPOT] before calling this routine
 431  * \param[in] inputrec  The input record
 432  */
 433 static void checkPotentialEnergyValidity(int64_t               step,
 434                                          const gmx_enerdata_t &enerd,
 435                                          const t_inputrec     &inputrec)
 436 {
 437     /* Threshold valid for comparing absolute potential energy against
 438      * the kinetic energy. Normally one should not consider absolute
 439      * potential energy values, but with a factor of one million
 440      * we should never get false positives.
 441      */
 442     constexpr real c_thresholdFactor = 1e6;
 443
 444     bool           energyIsNotFinite    = !std::isfinite(enerd.term[F_EPOT]);
 445     real           averageKineticEnergy = 0;
 446     /* We only check for large potential energy at the initial step,
 447      * because that is by far the most likely step for this too occur
 448      * and because computing the average kinetic energy is not free.
 449      * Note: nstcalcenergy >> 1 often does not allow to catch large energies
 450      * before they become NaN.
 451      */
 452     if (step == inputrec.init_step && EI_DYNAMICS(inputrec.eI))
 453     {
 454         averageKineticEnergy = averageKineticEnergyEstimate(inputrec.opts);
 455     }
 456
 457     if (energyIsNotFinite || (averageKineticEnergy > 0 &&
 458                               enerd.term[F_EPOT] > c_thresholdFactor*averageKineticEnergy))
 459     {
 460         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.",
 461                   step,
 462                   enerd.term[F_EPOT],
 463                   energyIsNotFinite ? "not finite" : "extremely high",
 464                   enerd.term[F_LJ],
 465                   enerd.term[F_COUL_SR],
 466                   energyIsNotFinite ? "non-finite" : "very high",
 467                   energyIsNotFinite ? " or Nan" : "");
 468     }
 469 }
 470
 471 /*! \brief Return true if there are special forces computed this step.
 472  *
 473  * The conditionals exactly correspond to those in computeSpecialForces().
 474  */
 475 static bool
 476 haveSpecialForces(const t_inputrec              *inputrec,
 477                   ForceProviders                *forceProviders,
 478                   const pull_t                  *pull_work,
 479                   int                            forceFlags,
 480                   const gmx_edsam               *ed)
 481 {
 482     const bool computeForces = (forceFlags & GMX_FORCE_FORCES) != 0;
 483
 484     return
 485         ((computeForces && forceProviders->hasForceProvider()) ||         // forceProviders
 486          (inputrec->bPull && pull_have_potential(pull_work)) ||           // pull
 487          inputrec->bRot ||                                                // enforced rotation
 488          (ed != nullptr) ||                                               // flooding
 489          (inputrec->bIMD && computeForces));                              // IMD
 490 }
 491
 492 /*! \brief Compute forces and/or energies for special algorithms
 493  *
 494  * The intention is to collect all calls to algorithms that compute
 495  * forces on local atoms only and that do not contribute to the local
 496  * virial sum (but add their virial contribution separately).
 497  * Eventually these should likely all become ForceProviders.
 498  * Within this function the intention is to have algorithms that do
 499  * global communication at the end, so global barriers within the MD loop
 500  * are as close together as possible.
 501  *
 502  * \param[in]     fplog            The log file
 503  * \param[in]     cr               The communication record
 504  * \param[in]     inputrec         The input record
 505  * \param[in]     awh              The Awh module (nullptr if none in use).
 506  * \param[in]     enforcedRotation Enforced rotation module.
 507  * \param[in]     imdSession       The IMD session
 508  * \param[in]     pull_work        The pull work structure.
 509  * \param[in]     step             The current MD step
 510  * \param[in]     t                The current time
 511  * \param[in,out] wcycle           Wallcycle accounting struct
 512  * \param[in,out] forceProviders   Pointer to a list of force providers
 513  * \param[in]     box              The unit cell
 514  * \param[in]     x                The coordinates
 515  * \param[in]     mdatoms          Per atom properties
 516  * \param[in]     lambda           Array of free-energy lambda values
 517  * \param[in]     forceFlags       Flags that tell whether we should compute forces/energies/virial
 518  * \param[in,out] forceWithVirial  Force and virial buffers
 519  * \param[in,out] enerd            Energy buffer
 520  * \param[in,out] ed               Essential dynamics pointer
 521  * \param[in]     bNS              Tells if we did neighbor searching this step, used for ED sampling
 522  *
 523  * \todo Remove bNS, which is used incorrectly.
 524  * \todo Convert all other algorithms called here to ForceProviders.
 525  */
 526 static void
 527 computeSpecialForces(FILE                          *fplog,
 528                      const t_commrec               *cr,
 529                      const t_inputrec              *inputrec,
 530                      gmx::Awh                      *awh,
 531                      gmx_enfrot                    *enforcedRotation,
 532                      gmx::ImdSession               *imdSession,
 533                      pull_t                        *pull_work,
 534                      int64_t                        step,
 535                      double                         t,
 536                      gmx_wallcycle_t                wcycle,
 537                      ForceProviders                *forceProviders,
 538                      matrix                         box,
 539                      gmx::ArrayRef<const gmx::RVec> x,
 540                      const t_mdatoms               *mdatoms,
 541                      real                          *lambda,
 542                      int                            forceFlags,
 543                      gmx::ForceWithVirial          *forceWithVirial,
 544                      gmx_enerdata_t                *enerd,
 545                      gmx_edsam                     *ed,
 546                      gmx_bool                       bNS)
 547 {
 548     const bool computeForces = (forceFlags & GMX_FORCE_FORCES) != 0;
 549
 550     /* NOTE: Currently all ForceProviders only provide forces.
 551      *       When they also provide energies, remove this conditional.
 552      */
 553     if (computeForces)
 554     {
 555         gmx::ForceProviderInput  forceProviderInput(x, *mdatoms, t, box, *cr);
 556         gmx::ForceProviderOutput forceProviderOutput(forceWithVirial, enerd);
 557
 558         /* Collect forces from modules */
 559         forceProviders->calculateForces(forceProviderInput, &forceProviderOutput);
 560     }
 561
 562     if (inputrec->bPull && pull_have_potential(pull_work))
 563     {
 564         pull_potential_wrapper(cr, inputrec, box, x,
 565                                forceWithVirial,
 566                                mdatoms, enerd, pull_work, lambda, t,
 567                                wcycle);
 568
 569         if (awh)
 570         {
 571             enerd->term[F_COM_PULL] +=
 572                 awh->applyBiasForcesAndUpdateBias(inputrec->ePBC, *mdatoms, box,
 573                                                   forceWithVirial,
 574                                                   t, step, wcycle, fplog);
 575         }
 576     }
 577
 578     rvec *f = as_rvec_array(forceWithVirial->force_.data());
 579
 580     /* Add the forces from enforced rotation potentials (if any) */
 581     if (inputrec->bRot)
 582     {
 583         wallcycle_start(wcycle, ewcROTadd);
 584         enerd->term[F_COM_PULL] += add_rot_forces(enforcedRotation, f, cr, step, t);
 585         wallcycle_stop(wcycle, ewcROTadd);
 586     }
 587
 588     if (ed)
 589     {
 590         /* Note that since init_edsam() is called after the initialization
 591          * of forcerec, edsam doesn't request the noVirSum force buffer.
 592          * Thus if no other algorithm (e.g. PME) requires it, the forces
 593          * here will contribute to the virial.
 594          */
 595         do_flood(cr, inputrec, as_rvec_array(x.data()), f, ed, box, step, bNS);
 596     }
 597
 598     /* Add forces from interactive molecular dynamics (IMD), if any */
 599     if (inputrec->bIMD && computeForces)
 600     {
 601         imdSession->applyForces(f);
 602     }
 603 }
 604
 605 /*! \brief Launch the prepare_step and spread stages of PME GPU.
 606  *
 607  * \param[in]  pmedata       The PME structure
 608  * \param[in]  box           The box matrix
 609  * \param[in]  x             Coordinate array
 610  * \param[in]  flags         Force flags
 611  * \param[in]  pmeFlags      PME flags
 612  * \param[in]  wcycle        The wallcycle structure
 613  */
 614 static inline void launchPmeGpuSpread(gmx_pme_t      *pmedata,
 615                                       matrix          box,
 616                                       rvec            x[],
 617                                       int             flags,
 618                                       int             pmeFlags,
 619                                       gmx_wallcycle_t wcycle)
 620 {
 621     pme_gpu_prepare_computation(pmedata, (flags & GMX_FORCE_DYNAMICBOX) != 0, box, wcycle, pmeFlags);
 622     pme_gpu_launch_spread(pmedata, x, wcycle);
 623 }
 624
 625 /*! \brief Launch the FFT and gather stages of PME GPU
 626  *
 627  * This function only implements setting the output forces (no accumulation).
 628  *
 629  * \param[in]  pmedata        The PME structure
 630  * \param[in]  wcycle         The wallcycle structure
 631  */
 632 static void launchPmeGpuFftAndGather(gmx_pme_t        *pmedata,
 633                                      gmx_wallcycle_t   wcycle)
 634 {
 635     pme_gpu_launch_complex_transforms(pmedata, wcycle);
 636     pme_gpu_launch_gather(pmedata, wcycle, PmeForceOutputHandling::Set);
 637 }
 638
 639 /*! \brief
 640  *  Polling wait for either of the PME or nonbonded GPU tasks.
 641  *
 642  * Instead of a static order in waiting for GPU tasks, this function
 643  * polls checking which of the two tasks completes first, and does the
 644  * associated force buffer reduction overlapped with the other task.
 645  * By doing that, unlike static scheduling order, it can always overlap
 646  * one of the reductions, regardless of the GPU task completion order.
 647  *
 648  * \param[in]     nbv              Nonbonded verlet structure
 649  * \param[in,out] pmedata          PME module data
 650  * \param[in,out] force            Force array to reduce task outputs into.
 651  * \param[in,out] forceWithVirial  Force and virial buffers
 652  * \param[in,out] fshift           Shift force output vector results are reduced into
 653  * \param[in,out] enerd            Energy data structure results are reduced into
 654  * \param[in]     flags            Force flags
 655  * \param[in]     pmeFlags         PME flags
 656  * \param[in]     haveOtherWork    Tells whether there is other work than non-bonded in the stream(s)
 657  * \param[in]     wcycle           The wallcycle structure
 658  */
 659 static void alternatePmeNbGpuWaitReduce(nonbonded_verlet_t                  *nbv,
 660                                         gmx_pme_t                           *pmedata,
 661                                         gmx::ArrayRefWithPadding<gmx::RVec> *force,
 662                                         gmx::ForceWithVirial                *forceWithVirial,
 663                                         rvec                                 fshift[],
 664                                         gmx_enerdata_t                      *enerd,
 665                                         int                                  flags,
 666                                         int                                  pmeFlags,
 667                                         bool                                 haveOtherWork,
 668                                         gmx_wallcycle_t                      wcycle)
 669 {
 670     bool isPmeGpuDone = false;
 671     bool isNbGpuDone  = false;
 672
 673
 674     gmx::ArrayRef<const gmx::RVec> pmeGpuForces;
 675
 676     while (!isPmeGpuDone || !isNbGpuDone)
 677     {
 678         if (!isPmeGpuDone)
 679         {
 680             GpuTaskCompletion completionType = (isNbGpuDone) ? GpuTaskCompletion::Wait : GpuTaskCompletion::Check;
 681             isPmeGpuDone = pme_gpu_try_finish_task(pmedata, pmeFlags, wcycle, forceWithVirial, enerd, completionType);
 682         }
 683
 684         if (!isNbGpuDone)
 685         {
 686             GpuTaskCompletion completionType = (isPmeGpuDone) ? GpuTaskCompletion::Wait : GpuTaskCompletion::Check;
 687             wallcycle_start_nocount(wcycle, ewcWAIT_GPU_NB_L);
 688             isNbGpuDone = Nbnxm::gpu_try_finish_task(nbv->gpu_nbv,
 689                                                      flags,
 690                                                      Nbnxm::AtomLocality::Local,
 691                                                      haveOtherWork,
 692                                                      enerd->grpp.ener[egLJSR].data(),
 693                                                      enerd->grpp.ener[egCOULSR].data(),
 694                                                      fshift, completionType);
 695             wallcycle_stop(wcycle, ewcWAIT_GPU_NB_L);
 696             // To get the call count right, when the task finished we
 697             // issue a start/stop.
 698             // TODO: move the ewcWAIT_GPU_NB_L cycle counting into nbnxn_gpu_try_finish_task()
 699             // and ewcNB_XF_BUF_OPS counting into nbnxn_atomdata_add_nbat_f_to_f().
 700             if (isNbGpuDone)
 701             {
 702                 wallcycle_start(wcycle, ewcWAIT_GPU_NB_L);
 703                 wallcycle_stop(wcycle, ewcWAIT_GPU_NB_L);
 704
 705                 nbv->atomdata_add_nbat_f_to_f(Nbnxm::AtomLocality::Local,
 706                                               as_rvec_array(force->unpaddedArrayRef().data()), wcycle);
 707             }
 708         }
 709     }
 710 }
 711
 712 /*! \brief Hack structure with force ouput buffers for do_force for the home atoms for this domain */
 713 struct ForceOutputs
 714 {
 715     //! Constructor
 716     ForceOutputs(rvec *f, gmx::ForceWithVirial const forceWithVirial) :
 717         f(f),
 718         forceWithVirial(forceWithVirial) {}
 719
 720     //! Force output buffer used by legacy modules (without SIMD padding)
 721     rvec                 *const f;
 722     //! Force with direct virial contribution (if there are any; without SIMD padding)
 723     gmx::ForceWithVirial        forceWithVirial;
 724 };
 725
 726 /*! \brief Set up the different force buffers; also does clearing.
 727  *
 728  * \param[in] fr        force record pointer
 729  * \param[in] pull_work The pull work object.
 730  * \param[in] inputrec  input record
 731  * \param[in] force     force array
 732  * \param[in] bDoForces True if force are computed this step
 733  * \param[in] doVirial  True if virial is computed this step
 734  * \param[out] wcycle   wallcycle recording structure
 735  *
 736  * \returns             Cleared force output structure
 737  */
 738 static ForceOutputs
 739 setupForceOutputs(const t_forcerec                    *fr,
 740                   pull_t                              *pull_work,
 741                   const t_inputrec                    &inputrec,
 742                   gmx::ArrayRefWithPadding<gmx::RVec>  force,
 743                   const bool                           bDoForces,
 744                   const bool                           doVirial,
 745                   gmx_wallcycle_t                      wcycle)
 746 {
 747     wallcycle_sub_start(wcycle, ewcsCLEAR_FORCE_BUFFER);
 748
 749     /* Temporary solution until all routines take PaddedRVecVector */
 750     rvec *const f = as_rvec_array(force.unpaddedArrayRef().data());
 751     if (bDoForces)
 752     {
 753         /* Clear the short- and long-range forces */
 754         clear_rvecs_omp(fr->natoms_force_constr, f);
 755     }
 756
 757     /* If we need to compute the virial, we might need a separate
 758      * force buffer for algorithms for which the virial is calculated
 759      * directly, such as PME. Otherwise, forceWithVirial uses the
 760      * the same force (f in legacy calls) buffer as other algorithms.
 761      */
 762     const bool useSeparateForceWithVirialBuffer = (bDoForces && (doVirial && fr->haveDirectVirialContributions));
 763
 764
 765     /* forceWithVirial uses the local atom range only */
 766     gmx::ForceWithVirial forceWithVirial (useSeparateForceWithVirialBuffer ?
 767                                           *fr->forceBufferForDirectVirialContributions : force.unpaddedArrayRef(),
 768                                           doVirial);
 769
 770     if (useSeparateForceWithVirialBuffer)
 771     {
 772         /* TODO: update comment
 773          * We only compute forces on local atoms. Note that vsites can
 774          * spread to non-local atoms, but that part of the buffer is
 775          * cleared separately in the vsite spreading code.
 776          */
 777         clear_rvecs_omp(forceWithVirial.force_.size(), as_rvec_array(forceWithVirial.force_.data()));
 778     }
 779
 780     if (inputrec.bPull && pull_have_constraint(pull_work))
 781     {
 782         clear_pull_forces(pull_work);
 783     }
 784
 785     wallcycle_sub_stop(wcycle, ewcsCLEAR_FORCE_BUFFER);
 786
 787     return ForceOutputs(f, forceWithVirial);
 788 }
 789
 790
 791 /*! \brief Set up flags that indicate what type of work is there to compute.
 792  *
 793  * Currently we only update it at search steps,
 794  * but some properties may change more frequently (e.g. virial/non-virial step),
 795  * so when including those either the frequency of update (per-step) or the scope
 796  * of a flag will change (i.e. a set of flags for nstlist steps).
 797  *
 798  */
 799 static void
 800 setupForceWorkload(gmx::PpForceWorkload *forceWork,
 801                    const t_inputrec     *inputrec,
 802                    const t_forcerec     *fr,
 803                    const pull_t         *pull_work,
 804                    const gmx_edsam      *ed,
 805                    const t_idef         &idef,
 806                    const t_fcdata       *fcd,
 807                    const int             forceFlags
 808                    )
 809 {
 810     forceWork->haveSpecialForces      = haveSpecialForces(inputrec, fr->forceProviders, pull_work, forceFlags, ed);
 811     forceWork->haveCpuBondedWork      = haveCpuBondeds(*fr);
 812     forceWork->haveGpuBondedWork      = ((fr->gpuBonded != nullptr) && fr->gpuBonded->haveInteractions());
 813     forceWork->haveRestraintsWork     = havePositionRestraints(idef, *fcd);
 814     forceWork->haveCpuListedForceWork = haveCpuListedForces(*fr, idef, *fcd);
 815 }
 816
 817 void do_force(FILE                                     *fplog,
 818               const t_commrec                          *cr,
 819               const gmx_multisim_t                     *ms,
 820               const t_inputrec                         *inputrec,
 821               gmx::Awh                                 *awh,
 822               gmx_enfrot                               *enforcedRotation,
 823               gmx::ImdSession                          *imdSession,
 824               pull_t                                   *pull_work,
 825               int64_t                                   step,
 826               t_nrnb                                   *nrnb,
 827               gmx_wallcycle_t                           wcycle,
 828               const gmx_localtop_t                     *top,
 829               matrix                                    box,
 830               gmx::ArrayRefWithPadding<gmx::RVec>       x,     //NOLINT(performance-unnecessary-value-param)
 831               history_t                                *hist,
 832               gmx::ArrayRefWithPadding<gmx::RVec>       force, //NOLINT(performance-unnecessary-value-param)
 833               tensor                                    vir_force,
 834               const t_mdatoms                          *mdatoms,
 835               gmx_enerdata_t                           *enerd,
 836               t_fcdata                                 *fcd,
 837               gmx::ArrayRef<real>                       lambda,
 838               t_graph                                  *graph,
 839               t_forcerec                               *fr,
 840               gmx::PpForceWorkload                     *ppForceWorkload,
 841               const gmx_vsite_t                        *vsite,
 842               rvec                                      mu_tot,
 843               double                                    t,
 844               gmx_edsam                                *ed,
 845               int                                       flags,
 846               const DDBalanceRegionHandler             &ddBalanceRegionHandler)
 847 {
 848     int                  i, j;
 849     double               mu[2*DIM];
 850     gmx_bool             bStateChanged, bNS, bFillGrid, bCalcCGCM;
 851     gmx_bool             bDoForces, bUseGPU, bUseOrEmulGPU;
 852     nonbonded_verlet_t  *nbv = fr->nbv.get();
 853     interaction_const_t *ic  = fr->ic;
 854
 855     /* modify force flag if not doing nonbonded */
 856     if (!fr->bNonbonded)
 857     {
 858         flags &= ~GMX_FORCE_NONBONDED;
 859     }
 860     bStateChanged = ((flags & GMX_FORCE_STATECHANGED) != 0);
 861     bNS           = ((flags & GMX_FORCE_NS) != 0);
 862     bFillGrid     = (bNS && bStateChanged);
 863     bCalcCGCM     = (bFillGrid && !DOMAINDECOMP(cr));
 864     bDoForces     = ((flags & GMX_FORCE_FORCES) != 0);
 865     bUseGPU       = fr->nbv->useGpu();
 866     bUseOrEmulGPU = bUseGPU || fr->nbv->emulateGpu();
 867
 868     const auto pmeRunMode = fr->pmedata ? pme_run_mode(fr->pmedata) : PmeRunMode::CPU;
 869     // TODO slim this conditional down - inputrec and duty checks should mean the same in proper code!
 870     const bool useGpuPme  = EEL_PME(fr->ic->eeltype) && thisRankHasDuty(cr, DUTY_PME) &&
 871         ((pmeRunMode == PmeRunMode::GPU) || (pmeRunMode == PmeRunMode::Mixed));
 872     const int  pmeFlags = GMX_PME_SPREAD | GMX_PME_SOLVE |
 873         ((flags & GMX_FORCE_VIRIAL) ? GMX_PME_CALC_ENER_VIR : 0) |
 874         ((flags & GMX_FORCE_ENERGY) ? GMX_PME_CALC_ENER_VIR : 0) |
 875         ((flags & GMX_FORCE_FORCES) ? GMX_PME_CALC_F : 0);
 876
 877     const bool useGpuXBufOps = (c_enableGpuBufOps && bUseGPU && (GMX_GPU == GMX_GPU_CUDA));
 878
 879     /* At a search step we need to start the first balancing region
 880      * somewhere early inside the step after communication during domain
 881      * decomposition (and not during the previous step as usual).
 882      */
 883     if (bNS)
 884     {
 885         ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::yes);
 886     }
 887
 888     const int start  = 0;
 889     const int homenr = mdatoms->homenr;
 890
 891     clear_mat(vir_force);
 892
 893     if (bStateChanged)
 894     {
 895         update_forcerec(fr, box);
 896
 897         if (inputrecNeedMutot(inputrec))
 898         {
 899             /* Calculate total (local) dipole moment in a temporary common array.
 900              * This makes it possible to sum them over nodes faster.
 901              */
 902             calc_mu(start, homenr,
 903                     x.unpaddedArrayRef(), mdatoms->chargeA, mdatoms->chargeB, mdatoms->nChargePerturbed,
 904                     mu, mu+DIM);
 905         }
 906     }
 907
 908     if (fr->ePBC != epbcNONE)
 909     {
 910         /* Compute shift vectors every step,
 911          * because of pressure coupling or box deformation!
 912          */
 913         if ((flags & GMX_FORCE_DYNAMICBOX) && bStateChanged)
 914         {
 915             calc_shifts(box, fr->shift_vec);
 916         }
 917
 918         if (bCalcCGCM)
 919         {
 920             put_atoms_in_box_omp(fr->ePBC, box, x.unpaddedArrayRef().subArray(0, homenr), gmx_omp_nthreads_get(emntDefault));
 921             inc_nrnb(nrnb, eNR_SHIFTX, homenr);
 922         }
 923         else if (EI_ENERGY_MINIMIZATION(inputrec->eI) && graph)
 924         {
 925             unshift_self(graph, box, as_rvec_array(x.unpaddedArrayRef().data()));
 926         }
 927     }
 928
 929     nbnxn_atomdata_copy_shiftvec((flags & GMX_FORCE_DYNAMICBOX) != 0,
 930                                  fr->shift_vec, nbv->nbat.get());
 931
 932 #if GMX_MPI
 933     if (!thisRankHasDuty(cr, DUTY_PME))
 934     {
 935         /* Send particle coordinates to the pme nodes.
 936          * Since this is only implemented for domain decomposition
 937          * and domain decomposition does not use the graph,
 938          * we do not need to worry about shifting.
 939          */
 940         gmx_pme_send_coordinates(cr, box, as_rvec_array(x.unpaddedArrayRef().data()),
 941                                  lambda[efptCOUL], lambda[efptVDW],
 942                                  (flags & (GMX_FORCE_VIRIAL | GMX_FORCE_ENERGY)) != 0,
 943                                  step, wcycle);
 944     }
 945 #endif /* GMX_MPI */
 946
 947     if (useGpuPme)
 948     {
 949         launchPmeGpuSpread(fr->pmedata, box, as_rvec_array(x.unpaddedArrayRef().data()), flags, pmeFlags, wcycle);
 950     }
 951
 952     /* do gridding for pair search */
 953     if (bNS)
 954     {
 955         if (graph && bStateChanged)
 956         {
 957             /* Calculate intramolecular shift vectors to make molecules whole */
 958             mk_mshift(fplog, graph, fr->ePBC, box, as_rvec_array(x.unpaddedArrayRef().data()));
 959         }
 960
 961         // TODO
 962         // - vzero is constant, do we need to pass it?
 963         // - box_diag should be passed directly to nbnxn_put_on_grid
 964         //
 965         rvec vzero;
 966         clear_rvec(vzero);
 967
 968         rvec box_diag;
 969         box_diag[XX] = box[XX][XX];
 970         box_diag[YY] = box[YY][YY];
 971         box_diag[ZZ] = box[ZZ][ZZ];
 972
 973         wallcycle_start(wcycle, ewcNS);
 974         if (!DOMAINDECOMP(cr))
 975         {
 976             wallcycle_sub_start(wcycle, ewcsNBS_GRID_LOCAL);
 977             nbnxn_put_on_grid(nbv, box,
 978                               0, vzero, box_diag,
 979                               nullptr, 0, mdatoms->homenr, -1,
 980                               fr->cginfo, x.unpaddedArrayRef(),
 981                               0, nullptr);
 982             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_LOCAL);
 983         }
 984         else
 985         {
 986             wallcycle_sub_start(wcycle, ewcsNBS_GRID_NONLOCAL);
 987             nbnxn_put_on_grid_nonlocal(nbv, domdec_zones(cr->dd),
 988                                        fr->cginfo, x.unpaddedArrayRef());
 989             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_NONLOCAL);
 990         }
 991
 992         nbv->setAtomProperties(*mdatoms, fr->cginfo);
 993
 994         wallcycle_stop(wcycle, ewcNS);
 995
 996         /* initialize the GPU nbnxm atom data and bonded data structures */
 997         if (bUseGPU)
 998         {
 999             wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1000
1001             wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1002             Nbnxm::gpu_init_atomdata(nbv->gpu_nbv, nbv->nbat.get());
1003             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1004
1005             if (fr->gpuBonded)
1006             {
1007                 /* Now we put all atoms on the grid, we can assign bonded
1008                  * interactions to the GPU, where the grid order is
1009                  * needed. Also the xq, f and fshift device buffers have
1010                  * been reallocated if needed, so the bonded code can
1011                  * learn about them. */
1012                 // TODO the xq, f, and fshift buffers are now shared
1013                 // resources, so they should be maintained by a
1014                 // higher-level object than the nb module.
1015                 fr->gpuBonded->updateInteractionListsAndDeviceBuffers(nbv->getGridIndices(),
1016                                                                       top->idef,
1017                                                                       Nbnxm::gpu_get_xq(nbv->gpu_nbv),
1018                                                                       Nbnxm::gpu_get_f(nbv->gpu_nbv),
1019                                                                       Nbnxm::gpu_get_fshift(nbv->gpu_nbv));
1020             }
1021             wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1022         }
1023
1024         // Need to run after the GPU-offload bonded interaction lists
1025         // are set up to be able to determine whether there is bonded work.
1026         setupForceWorkload(ppForceWorkload,
1027                            inputrec,
1028                            fr,
1029                            pull_work,
1030                            ed,
1031                            top->idef,
1032                            fcd,
1033                            flags);
1034     }
1035
1036     /* do local pair search */
1037     if (bNS)
1038     {
1039         // TODO: fuse this branch with the above bNS block
1040         wallcycle_start_nocount(wcycle, ewcNS);
1041         wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_LOCAL);
1042         /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
1043         nbv->constructPairlist(Nbnxm::InteractionLocality::Local,
1044                                &top->excls, step, nrnb);
1045         wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_LOCAL);
1046         wallcycle_stop(wcycle, ewcNS);
1047
1048         if (useGpuXBufOps)
1049         {
1050             nbv->atomdata_init_copy_x_to_nbat_x_gpu( Nbnxm::AtomLocality::Local);
1051         }
1052
1053     }
1054     else
1055     {
1056         nbv->setCoordinates(Nbnxm::AtomLocality::Local, false,
1057                             x.unpaddedArrayRef(), useGpuXBufOps, pme_gpu_get_device_x(fr->pmedata), wcycle);
1058     }
1059
1060     if (bUseGPU)
1061     {
1062         ddBalanceRegionHandler.openBeforeForceComputationGpu();
1063
1064         wallcycle_start(wcycle, ewcLAUNCH_GPU);
1065
1066         wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1067         Nbnxm::gpu_upload_shiftvec(nbv->gpu_nbv, nbv->nbat.get());
1068         if (bNS || !useGpuXBufOps)
1069         {
1070             Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(),
1071                                       Nbnxm::AtomLocality::Local,
1072                                       ppForceWorkload->haveGpuBondedWork);
1073         }
1074         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1075         // with X buffer ops offloaded to the GPU on all but the search steps
1076
1077         // bonded work not split into separate local and non-local, so with DD
1078         // we can only launch the kernel after non-local coordinates have been received.
1079         if (ppForceWorkload->haveGpuBondedWork && !havePPDomainDecomposition(cr))
1080         {
1081             wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_BONDED);
1082             fr->gpuBonded->launchKernel(fr, flags, box);
1083             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1084         }
1085
1086         /* launch local nonbonded work on GPU */
1087         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1088         do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::Local, enbvClearFNo,
1089                      step, nrnb, wcycle);
1090         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1091         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1092     }
1093
1094     if (useGpuPme)
1095     {
1096         // In PME GPU and mixed mode we launch FFT / gather after the
1097         // X copy/transform to allow overlap as well as after the GPU NB
1098         // launch to avoid FFT launch overhead hijacking the CPU and delaying
1099         // the nonbonded kernel.
1100         launchPmeGpuFftAndGather(fr->pmedata, wcycle);
1101     }
1102
1103     /* Communicate coordinates and sum dipole if necessary +
1104        do non-local pair search */
1105     if (havePPDomainDecomposition(cr))
1106     {
1107         if (bNS)
1108         {
1109             // TODO: fuse this branch with the above large bNS block
1110             wallcycle_start_nocount(wcycle, ewcNS);
1111             wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1112             /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
1113             nbv->constructPairlist(Nbnxm::InteractionLocality::NonLocal,
1114                                    &top->excls, step, nrnb);
1115             wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1116             wallcycle_stop(wcycle, ewcNS);
1117
1118             if (useGpuXBufOps)
1119             {
1120
1121                 nbv->atomdata_init_copy_x_to_nbat_x_gpu( Nbnxm::AtomLocality::NonLocal);
1122             }
1123         }
1124         else
1125         {
1126             dd_move_x(cr->dd, box, x.unpaddedArrayRef(), wcycle);
1127
1128             nbv->setCoordinates(Nbnxm::AtomLocality::NonLocal, false,
1129                                 x.unpaddedArrayRef(), useGpuXBufOps, pme_gpu_get_device_x(fr->pmedata), wcycle);
1130
1131         }
1132
1133         if (bUseGPU)
1134         {
1135             wallcycle_start(wcycle, ewcLAUNCH_GPU);
1136
1137             if (bNS || !useGpuXBufOps)
1138             {
1139                 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1140                 Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(),
1141                                           Nbnxm::AtomLocality::NonLocal,
1142                                           ppForceWorkload->haveGpuBondedWork);
1143                 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1144             }
1145
1146             if (ppForceWorkload->haveGpuBondedWork)
1147             {
1148                 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_BONDED);
1149                 fr->gpuBonded->launchKernel(fr, flags, box);
1150                 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1151             }
1152
1153             /* launch non-local nonbonded tasks on GPU */
1154             wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1155             do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::NonLocal, enbvClearFNo,
1156                          step, nrnb, wcycle);
1157             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1158
1159             wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1160         }
1161     }
1162
1163     if (bUseGPU)
1164     {
1165         /* launch D2H copy-back F */
1166         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1167         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1168         if (havePPDomainDecomposition(cr))
1169         {
1170             Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat.get(),
1171                                       flags, Nbnxm::AtomLocality::NonLocal, ppForceWorkload->haveGpuBondedWork);
1172         }
1173         Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat.get(),
1174                                   flags, Nbnxm::AtomLocality::Local, ppForceWorkload->haveGpuBondedWork);
1175         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1176
1177         if (ppForceWorkload->haveGpuBondedWork && (flags & GMX_FORCE_ENERGY))
1178         {
1179             wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_BONDED);
1180             fr->gpuBonded->launchEnergyTransfer();
1181             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1182         }
1183         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1184     }
1185
1186     if (bStateChanged && inputrecNeedMutot(inputrec))
1187     {
1188         if (PAR(cr))
1189         {
1190             gmx_sumd(2*DIM, mu, cr);
1191
1192             ddBalanceRegionHandler.reopenRegionCpu();
1193         }
1194
1195         for (i = 0; i < 2; i++)
1196         {
1197             for (j = 0; j < DIM; j++)
1198             {
1199                 fr->mu_tot[i][j] = mu[i*DIM + j];
1200             }
1201         }
1202     }
1203     if (fr->efep == efepNO)
1204     {
1205         copy_rvec(fr->mu_tot[0], mu_tot);
1206     }
1207     else
1208     {
1209         for (j = 0; j < DIM; j++)
1210         {
1211             mu_tot[j] =
1212                 (1.0 - lambda[efptCOUL])*fr->mu_tot[0][j] +
1213                 lambda[efptCOUL]*fr->mu_tot[1][j];
1214         }
1215     }
1216
1217     /* Reset energies */
1218     reset_enerdata(enerd);
1219     clear_rvecs(SHIFTS, fr->fshift);
1220
1221     if (DOMAINDECOMP(cr) && !thisRankHasDuty(cr, DUTY_PME))
1222     {
1223         wallcycle_start(wcycle, ewcPPDURINGPME);
1224         dd_force_flop_start(cr->dd, nrnb);
1225     }
1226
1227     if (inputrec->bRot)
1228     {
1229         wallcycle_start(wcycle, ewcROT);
1230         do_rotation(cr, enforcedRotation, box, as_rvec_array(x.unpaddedArrayRef().data()), t, step, bNS);
1231         wallcycle_stop(wcycle, ewcROT);
1232     }
1233
1234     /* Start the force cycle counter.
1235      * Note that a different counter is used for dynamic load balancing.
1236      */
1237     wallcycle_start(wcycle, ewcFORCE);
1238
1239     // set up and clear force outputs
1240     struct ForceOutputs forceOut = setupForceOutputs(fr, pull_work, *inputrec, force, bDoForces,
1241                                                      ((flags & GMX_FORCE_VIRIAL) != 0), wcycle);
1242
1243     /* We calculate the non-bonded forces, when done on the CPU, here.
1244      * We do this before calling do_force_lowlevel, because in that
1245      * function, the listed forces are calculated before PME, which
1246      * does communication.  With this order, non-bonded and listed
1247      * force calculation imbalance can be balanced out by the domain
1248      * decomposition load balancing.
1249      */
1250
1251     if (!bUseOrEmulGPU)
1252     {
1253         do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::Local, enbvClearFYes,
1254                      step, nrnb, wcycle);
1255     }
1256
1257     if (fr->efep != efepNO)
1258     {
1259         /* Calculate the local and non-local free energy interactions here.
1260          * Happens here on the CPU both with and without GPU.
1261          */
1262         wallcycle_sub_start(wcycle, ewcsNONBONDED);
1263         nbv->dispatchFreeEnergyKernel(Nbnxm::InteractionLocality::Local,
1264                                       fr, as_rvec_array(x.unpaddedArrayRef().data()), forceOut.f, *mdatoms,
1265                                       inputrec->fepvals, lambda.data(),
1266                                       enerd, flags, nrnb);
1267
1268         if (havePPDomainDecomposition(cr))
1269         {
1270             nbv->dispatchFreeEnergyKernel(Nbnxm::InteractionLocality::NonLocal,
1271                                           fr, as_rvec_array(x.unpaddedArrayRef().data()), forceOut.f, *mdatoms,
1272                                           inputrec->fepvals, lambda.data(),
1273                                           enerd, flags, nrnb);
1274         }
1275         wallcycle_sub_stop(wcycle, ewcsNONBONDED);
1276     }
1277
1278     if (!bUseOrEmulGPU)
1279     {
1280         if (havePPDomainDecomposition(cr))
1281         {
1282             do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::NonLocal, enbvClearFNo,
1283                          step, nrnb, wcycle);
1284         }
1285
1286         /* Add all the non-bonded force to the normal force array.
1287          * This can be split into a local and a non-local part when overlapping
1288          * communication with calculation with domain decomposition.
1289          */
1290         wallcycle_stop(wcycle, ewcFORCE);
1291
1292         nbv->atomdata_add_nbat_f_to_f(Nbnxm::AtomLocality::All, forceOut.f, wcycle);
1293
1294         wallcycle_start_nocount(wcycle, ewcFORCE);
1295
1296         /* If there are multiple fshift output buffers we need to reduce them */
1297         if (flags & GMX_FORCE_VIRIAL)
1298         {
1299             /* This is not in a subcounter because it takes a
1300                negligible and constant-sized amount of time */
1301             nbnxn_atomdata_add_nbat_fshift_to_fshift(nbv->nbat.get(),
1302                                                      fr->fshift);
1303         }
1304     }
1305
1306     /* update QMMMrec, if necessary */
1307     if (fr->bQMMM)
1308     {
1309         update_QMMMrec(cr, fr, as_rvec_array(x.unpaddedArrayRef().data()), mdatoms, box);
1310     }
1311
1312     /* Compute the bonded and non-bonded energies and optionally forces */
1313     do_force_lowlevel(fr, inputrec, &(top->idef),
1314                       cr, ms, nrnb, wcycle, mdatoms,
1315                       x, hist, forceOut.f, &forceOut.forceWithVirial, enerd, fcd,
1316                       box, lambda.data(), graph, fr->mu_tot,
1317                       flags,
1318                       ddBalanceRegionHandler);
1319
1320     wallcycle_stop(wcycle, ewcFORCE);
1321
1322     computeSpecialForces(fplog, cr, inputrec, awh, enforcedRotation,
1323                          imdSession, pull_work, step, t, wcycle,
1324                          fr->forceProviders, box, x.unpaddedArrayRef(), mdatoms, lambda.data(),
1325                          flags, &forceOut.forceWithVirial, enerd,
1326                          ed, bNS);
1327
1328     // Will store the amount of cycles spent waiting for the GPU that
1329     // will be later used in the DLB accounting.
1330     float cycles_wait_gpu = 0;
1331     if (bUseOrEmulGPU)
1332     {
1333         /* wait for non-local forces (or calculate in emulation mode) */
1334         if (havePPDomainDecomposition(cr))
1335         {
1336             if (bUseGPU)
1337             {
1338                 wallcycle_start(wcycle, ewcWAIT_GPU_NB_NL);
1339                 Nbnxm::gpu_wait_finish_task(nbv->gpu_nbv,
1340                                             flags, Nbnxm::AtomLocality::NonLocal,
1341                                             ppForceWorkload->haveGpuBondedWork,
1342                                             enerd->grpp.ener[egLJSR].data(),
1343                                             enerd->grpp.ener[egCOULSR].data(),
1344                                             fr->fshift);
1345                 cycles_wait_gpu += wallcycle_stop(wcycle, ewcWAIT_GPU_NB_NL);
1346             }
1347             else
1348             {
1349                 wallcycle_start_nocount(wcycle, ewcFORCE);
1350                 do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::NonLocal, enbvClearFYes,
1351                              step, nrnb, wcycle);
1352                 wallcycle_stop(wcycle, ewcFORCE);
1353             }
1354
1355             nbv->atomdata_add_nbat_f_to_f(Nbnxm::AtomLocality::NonLocal,
1356                                           forceOut.f, wcycle);
1357
1358             if (fr->nbv->emulateGpu() && (flags & GMX_FORCE_VIRIAL))
1359             {
1360                 nbnxn_atomdata_add_nbat_fshift_to_fshift(nbv->nbat.get(),
1361                                                          fr->fshift);
1362             }
1363         }
1364     }
1365
1366     if (havePPDomainDecomposition(cr))
1367     {
1368         /* We are done with the CPU compute.
1369          * We will now communicate the non-local forces.
1370          * If we use a GPU this will overlap with GPU work, so in that case
1371          * we do not close the DD force balancing region here.
1372          */
1373         ddBalanceRegionHandler.closeAfterForceComputationCpu();
1374
1375         if (bDoForces)
1376         {
1377             dd_move_f(cr->dd, force.unpaddedArrayRef(), fr->fshift, wcycle);
1378         }
1379     }
1380
1381     // With both nonbonded and PME offloaded a GPU on the same rank, we use
1382     // an alternating wait/reduction scheme.
1383     bool alternateGpuWait = (!c_disableAlternatingWait && useGpuPme && bUseGPU && !DOMAINDECOMP(cr));
1384     if (alternateGpuWait)
1385     {
1386         alternatePmeNbGpuWaitReduce(fr->nbv.get(), fr->pmedata, &force, &forceOut.forceWithVirial, fr->fshift, enerd,
1387                                     flags, pmeFlags, ppForceWorkload->haveGpuBondedWork, wcycle);
1388     }
1389
1390     if (!alternateGpuWait && useGpuPme)
1391     {
1392         pme_gpu_wait_and_reduce(fr->pmedata, pmeFlags, wcycle, &forceOut.forceWithVirial, enerd);
1393     }
1394
1395     /* Wait for local GPU NB outputs on the non-alternating wait path */
1396     if (!alternateGpuWait && bUseGPU)
1397     {
1398         /* Measured overhead on CUDA and OpenCL with(out) GPU sharing
1399          * is between 0.5 and 1.5 Mcycles. So 2 MCycles is an overestimate,
1400          * but even with a step of 0.1 ms the difference is less than 1%
1401          * of the step time.
1402          */
1403         const float gpuWaitApiOverheadMargin = 2e6f; /* cycles */
1404
1405         wallcycle_start(wcycle, ewcWAIT_GPU_NB_L);
1406         Nbnxm::gpu_wait_finish_task(nbv->gpu_nbv,
1407                                     flags, Nbnxm::AtomLocality::Local, ppForceWorkload->haveGpuBondedWork,
1408                                     enerd->grpp.ener[egLJSR].data(),
1409                                     enerd->grpp.ener[egCOULSR].data(),
1410                                     fr->fshift);
1411         float cycles_tmp = wallcycle_stop(wcycle, ewcWAIT_GPU_NB_L);
1412
1413         if (ddBalanceRegionHandler.useBalancingRegion())
1414         {
1415             DdBalanceRegionWaitedForGpu waitedForGpu = DdBalanceRegionWaitedForGpu::yes;
1416             if (bDoForces && cycles_tmp <= gpuWaitApiOverheadMargin)
1417             {
1418                 /* We measured few cycles, it could be that the kernel
1419                  * and transfer finished earlier and there was no actual
1420                  * wait time, only API call overhead.
1421                  * Then the actual time could be anywhere between 0 and
1422                  * cycles_wait_est. We will use half of cycles_wait_est.
1423                  */
1424                 waitedForGpu = DdBalanceRegionWaitedForGpu::no;
1425             }
1426             ddBalanceRegionHandler.closeAfterForceComputationGpu(cycles_wait_gpu, waitedForGpu);
1427         }
1428     }
1429
1430     if (fr->nbv->emulateGpu())
1431     {
1432         // NOTE: emulation kernel is not included in the balancing region,
1433         // but emulation mode does not target performance anyway
1434         wallcycle_start_nocount(wcycle, ewcFORCE);
1435         do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::Local,
1436                      DOMAINDECOMP(cr) ? enbvClearFNo : enbvClearFYes,
1437                      step, nrnb, wcycle);
1438         wallcycle_stop(wcycle, ewcFORCE);
1439     }
1440
1441     if (useGpuPme)
1442     {
1443         pme_gpu_reinit_computation(fr->pmedata, wcycle);
1444     }
1445
1446     if (bUseGPU)
1447     {
1448         /* now clear the GPU outputs while we finish the step on the CPU */
1449         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1450         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1451         Nbnxm::gpu_clear_outputs(nbv->gpu_nbv, flags);
1452
1453         if (nbv->isDynamicPruningStepGpu(step))
1454         {
1455             nbv->dispatchPruneKernelGpu(step);
1456         }
1457         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1458         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1459     }
1460
1461     if (ppForceWorkload->haveGpuBondedWork && (flags & GMX_FORCE_ENERGY))
1462     {
1463         wallcycle_start(wcycle, ewcWAIT_GPU_BONDED);
1464         // in principle this should be included in the DD balancing region,
1465         // but generally it is infrequent so we'll omit it for the sake of
1466         // simpler code
1467         fr->gpuBonded->accumulateEnergyTerms(enerd);
1468         wallcycle_stop(wcycle, ewcWAIT_GPU_BONDED);
1469
1470         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1471         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_BONDED);
1472         fr->gpuBonded->clearEnergies();
1473         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1474         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1475     }
1476
1477     /* Do the nonbonded GPU (or emulation) force buffer reduction
1478      * on the non-alternating path. */
1479     if (bUseOrEmulGPU && !alternateGpuWait)
1480     {
1481         nbv->atomdata_add_nbat_f_to_f(Nbnxm::AtomLocality::Local,
1482                                       forceOut.f, wcycle);
1483     }
1484     if (DOMAINDECOMP(cr))
1485     {
1486         dd_force_flop_stop(cr->dd, nrnb);
1487     }
1488
1489     if (bDoForces)
1490     {
1491         /* If we have NoVirSum forces, but we do not calculate the virial,
1492          * we sum fr->f_novirsum=forceOut.f later.
1493          */
1494         if (vsite && !(fr->haveDirectVirialContributions && !(flags & GMX_FORCE_VIRIAL)))
1495         {
1496             spread_vsite_f(vsite, as_rvec_array(x.unpaddedArrayRef().data()), forceOut.f, fr->fshift, FALSE, nullptr, nrnb,
1497                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr, wcycle);
1498         }
1499
1500         if (flags & GMX_FORCE_VIRIAL)
1501         {
1502             /* Calculation of the virial must be done after vsites! */
1503             calc_virial(0, mdatoms->homenr, as_rvec_array(x.unpaddedArrayRef().data()), forceOut.f,
1504                         vir_force, graph, box, nrnb, fr, inputrec->ePBC);
1505         }
1506     }
1507
1508     if (PAR(cr) && !thisRankHasDuty(cr, DUTY_PME))
1509     {
1510         /* In case of node-splitting, the PP nodes receive the long-range
1511          * forces, virial and energy from the PME nodes here.
1512          */
1513         pme_receive_force_ener(cr, &forceOut.forceWithVirial, enerd, wcycle);
1514     }
1515
1516     if (bDoForces)
1517     {
1518         post_process_forces(cr, step, nrnb, wcycle,
1519                             top, box, as_rvec_array(x.unpaddedArrayRef().data()), forceOut.f, &forceOut.forceWithVirial,
1520                             vir_force, mdatoms, graph, fr, vsite,
1521                             flags);
1522     }
1523
1524     if (flags & GMX_FORCE_ENERGY)
1525     {
1526         /* Sum the potential energy terms from group contributions */
1527         sum_epot(&(enerd->grpp), enerd->term);
1528
1529         if (!EI_TPI(inputrec->eI))
1530         {
1531             checkPotentialEnergyValidity(step, *enerd, *inputrec);
1532         }
1533     }
1534
1535     /* In case we don't have constraints and are using GPUs, the next balancing
1536      * region starts here.
1537      * Some "special" work at the end of do_force_cuts?, such as vsite spread,
1538      * virial calculation and COM pulling, is not thus not included in
1539      * the balance timing, which is ok as most tasks do communication.
1540      */
1541     ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::no);
1542 }