2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
5 * Copyright (c) 2001-2004, The GROMACS development team.
6 * Copyright (c) 2013-2019,2020, 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.
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.
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.
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.
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.
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.
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/gpuhaloexchange.h"
53 #include "gromacs/domdec/partition.h"
54 #include "gromacs/essentialdynamics/edsam.h"
55 #include "gromacs/ewald/pme.h"
56 #include "gromacs/ewald/pme_pp.h"
57 #include "gromacs/ewald/pme_pp_comm_gpu.h"
58 #include "gromacs/gmxlib/network.h"
59 #include "gromacs/gmxlib/nonbonded/nb_free_energy.h"
60 #include "gromacs/gmxlib/nonbonded/nb_kernel.h"
61 #include "gromacs/gmxlib/nonbonded/nonbonded.h"
62 #include "gromacs/gpu_utils/gpu_utils.h"
63 #include "gromacs/imd/imd.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/orires.h"
68 #include "gromacs/math/arrayrefwithpadding.h"
69 #include "gromacs/math/functions.h"
70 #include "gromacs/math/units.h"
71 #include "gromacs/math/vec.h"
72 #include "gromacs/math/vecdump.h"
73 #include "gromacs/mdlib/calcmu.h"
74 #include "gromacs/mdlib/calcvir.h"
75 #include "gromacs/mdlib/constr.h"
76 #include "gromacs/mdlib/enerdata_utils.h"
77 #include "gromacs/mdlib/force.h"
78 #include "gromacs/mdlib/force_flags.h"
79 #include "gromacs/mdlib/forcerec.h"
80 #include "gromacs/mdlib/gmx_omp_nthreads.h"
81 #include "gromacs/mdlib/update.h"
82 #include "gromacs/mdlib/vsite.h"
83 #include "gromacs/mdlib/wholemoleculetransform.h"
84 #include "gromacs/mdtypes/commrec.h"
85 #include "gromacs/mdtypes/enerdata.h"
86 #include "gromacs/mdtypes/forceoutput.h"
87 #include "gromacs/mdtypes/forcerec.h"
88 #include "gromacs/mdtypes/iforceprovider.h"
89 #include "gromacs/mdtypes/inputrec.h"
90 #include "gromacs/mdtypes/md_enums.h"
91 #include "gromacs/mdtypes/mdatom.h"
92 #include "gromacs/mdtypes/simulation_workload.h"
93 #include "gromacs/mdtypes/state.h"
94 #include "gromacs/mdtypes/state_propagator_data_gpu.h"
95 #include "gromacs/nbnxm/gpu_data_mgmt.h"
96 #include "gromacs/nbnxm/nbnxm.h"
97 #include "gromacs/nbnxm/nbnxm_gpu.h"
98 #include "gromacs/pbcutil/ishift.h"
99 #include "gromacs/pbcutil/pbc.h"
100 #include "gromacs/pulling/pull.h"
101 #include "gromacs/pulling/pull_rotation.h"
102 #include "gromacs/timing/cyclecounter.h"
103 #include "gromacs/timing/gpu_timing.h"
104 #include "gromacs/timing/wallcycle.h"
105 #include "gromacs/timing/wallcyclereporting.h"
106 #include "gromacs/timing/walltime_accounting.h"
107 #include "gromacs/topology/topology.h"
108 #include "gromacs/utility/arrayref.h"
109 #include "gromacs/utility/basedefinitions.h"
110 #include "gromacs/utility/cstringutil.h"
111 #include "gromacs/utility/exceptions.h"
112 #include "gromacs/utility/fatalerror.h"
113 #include "gromacs/utility/fixedcapacityvector.h"
114 #include "gromacs/utility/gmxassert.h"
115 #include "gromacs/utility/gmxmpi.h"
116 #include "gromacs/utility/logger.h"
117 #include "gromacs/utility/smalloc.h"
118 #include "gromacs/utility/strconvert.h"
119 #include "gromacs/utility/sysinfo.h"
122 using gmx::AtomLocality;
123 using gmx::DomainLifetimeWorkload;
124 using gmx::ForceOutputs;
125 using gmx::ForceWithShiftForces;
126 using gmx::InteractionLocality;
128 using gmx::SimulationWorkload;
129 using gmx::StepWorkload;
131 // TODO: this environment variable allows us to verify before release
132 // that on less common architectures the total cost of polling is not larger than
133 // a blocking wait (so polling does not introduce overhead when the static
134 // PME-first ordering would suffice).
135 static const bool c_disableAlternatingWait = (getenv("GMX_DISABLE_ALTERNATING_GPU_WAIT") != nullptr);
137 static void sum_forces(ArrayRef<RVec> f, ArrayRef<const RVec> forceToAdd)
139 GMX_ASSERT(f.size() >= forceToAdd.size(), "Accumulation buffer should be sufficiently large");
140 const int end = forceToAdd.size();
142 int gmx_unused nt = gmx_omp_nthreads_get(emntDefault);
143 #pragma omp parallel for num_threads(nt) schedule(static)
144 for (int i = 0; i < end; i++)
146 rvec_inc(f[i], forceToAdd[i]);
150 static void calc_virial(int start,
153 const gmx::ForceWithShiftForces& forceWithShiftForces,
157 const t_forcerec* fr,
160 /* The short-range virial from surrounding boxes */
161 const rvec* fshift = as_rvec_array(forceWithShiftForces.shiftForces().data());
162 calc_vir(SHIFTS, fr->shift_vec, fshift, vir_part, pbcType == PbcType::Screw, box);
163 inc_nrnb(nrnb, eNR_VIRIAL, SHIFTS);
165 /* Calculate partial virial, for local atoms only, based on short range.
166 * Total virial is computed in global_stat, called from do_md
168 const rvec* f = as_rvec_array(forceWithShiftForces.force().data());
169 f_calc_vir(start, start + homenr, x, f, vir_part, box);
170 inc_nrnb(nrnb, eNR_VIRIAL, homenr);
174 pr_rvecs(debug, 0, "vir_part", vir_part, DIM);
178 static void pull_potential_wrapper(const t_commrec* cr,
179 const t_inputrec* ir,
181 gmx::ArrayRef<const gmx::RVec> x,
182 gmx::ForceWithVirial* force,
183 const t_mdatoms* mdatoms,
184 gmx_enerdata_t* enerd,
188 gmx_wallcycle_t wcycle)
193 /* Calculate the center of mass forces, this requires communication,
194 * which is why pull_potential is called close to other communication.
196 wallcycle_start(wcycle, ewcPULLPOT);
197 set_pbc(&pbc, ir->pbcType, box);
199 enerd->term[F_COM_PULL] +=
200 pull_potential(pull_work, mdatoms->massT, &pbc, cr, t, lambda[efptRESTRAINT],
201 as_rvec_array(x.data()), force, &dvdl);
202 enerd->dvdl_lin[efptRESTRAINT] += dvdl;
203 for (auto& dhdl : enerd->dhdlLambda)
207 wallcycle_stop(wcycle, ewcPULLPOT);
210 static void pme_receive_force_ener(t_forcerec* fr,
212 gmx::ForceWithVirial* forceWithVirial,
213 gmx_enerdata_t* enerd,
214 bool useGpuPmePpComms,
215 bool receivePmeForceToGpu,
216 gmx_wallcycle_t wcycle)
218 real e_q, e_lj, dvdl_q, dvdl_lj;
219 float cycles_ppdpme, cycles_seppme;
221 cycles_ppdpme = wallcycle_stop(wcycle, ewcPPDURINGPME);
222 dd_cycles_add(cr->dd, cycles_ppdpme, ddCyclPPduringPME);
224 /* In case of node-splitting, the PP nodes receive the long-range
225 * forces, virial and energy from the PME nodes here.
227 wallcycle_start(wcycle, ewcPP_PMEWAITRECVF);
230 gmx_pme_receive_f(fr->pmePpCommGpu.get(), cr, forceWithVirial, &e_q, &e_lj, &dvdl_q, &dvdl_lj,
231 useGpuPmePpComms, receivePmeForceToGpu, &cycles_seppme);
232 enerd->term[F_COUL_RECIP] += e_q;
233 enerd->term[F_LJ_RECIP] += e_lj;
234 enerd->dvdl_lin[efptCOUL] += dvdl_q;
235 enerd->dvdl_lin[efptVDW] += dvdl_lj;
237 for (auto& dhdl : enerd->dhdlLambda)
239 dhdl += dvdl_q + dvdl_lj;
244 dd_cycles_add(cr->dd, cycles_seppme, ddCyclPME);
246 wallcycle_stop(wcycle, ewcPP_PMEWAITRECVF);
249 static void print_large_forces(FILE* fp,
254 ArrayRef<const RVec> x,
255 ArrayRef<const RVec> f)
257 real force2Tolerance = gmx::square(forceTolerance);
258 gmx::index numNonFinite = 0;
259 for (int i = 0; i < md->homenr; i++)
261 real force2 = norm2(f[i]);
262 bool nonFinite = !std::isfinite(force2);
263 if (force2 >= force2Tolerance || nonFinite)
265 fprintf(fp, "step %" PRId64 " atom %6d x %8.3f %8.3f %8.3f force %12.5e\n", step,
266 ddglatnr(cr->dd, i), x[i][XX], x[i][YY], x[i][ZZ], std::sqrt(force2));
273 if (numNonFinite > 0)
275 /* Note that with MPI this fatal call on one rank might interrupt
276 * the printing on other ranks. But we can only avoid that with
277 * an expensive MPI barrier that we would need at each step.
279 gmx_fatal(FARGS, "At step %" PRId64 " detected non-finite forces on %td atoms", step, numNonFinite);
283 //! When necessary, spreads forces on vsites and computes the virial for \p forceOutputs->forceWithShiftForces()
284 static void postProcessForceWithShiftForces(t_nrnb* nrnb,
285 gmx_wallcycle_t wcycle,
287 ArrayRef<const RVec> x,
288 ForceOutputs* forceOutputs,
290 const t_mdatoms& mdatoms,
291 const t_forcerec& fr,
292 gmx::VirtualSitesHandler* vsite,
293 const StepWorkload& stepWork)
295 ForceWithShiftForces& forceWithShiftForces = forceOutputs->forceWithShiftForces();
297 /* If we have NoVirSum forces, but we do not calculate the virial,
298 * we later sum the forceWithShiftForces buffer together with
299 * the noVirSum buffer and spread the combined vsite forces at once.
301 if (vsite && (!forceOutputs->haveForceWithVirial() || stepWork.computeVirial))
303 using VirialHandling = gmx::VirtualSitesHandler::VirialHandling;
305 auto f = forceWithShiftForces.force();
306 auto fshift = forceWithShiftForces.shiftForces();
307 const VirialHandling virialHandling =
308 (stepWork.computeVirial ? VirialHandling::Pbc : VirialHandling::None);
309 vsite->spreadForces(x, f, virialHandling, fshift, nullptr, nrnb, box, wcycle);
310 forceWithShiftForces.haveSpreadVsiteForces() = true;
313 if (stepWork.computeVirial)
315 /* Calculation of the virial must be done after vsites! */
316 calc_virial(0, mdatoms.homenr, as_rvec_array(x.data()), forceWithShiftForces, vir_force,
317 box, nrnb, &fr, fr.pbcType);
321 //! Spread, compute virial for and sum forces, when necessary
322 static void postProcessForces(const t_commrec* cr,
325 gmx_wallcycle_t wcycle,
327 ArrayRef<const RVec> x,
328 ForceOutputs* forceOutputs,
330 const t_mdatoms* mdatoms,
331 const t_forcerec* fr,
332 gmx::VirtualSitesHandler* vsite,
333 const StepWorkload& stepWork)
335 // Extract the final output force buffer, which is also the buffer for forces with shift forces
336 ArrayRef<RVec> f = forceOutputs->forceWithShiftForces().force();
338 if (forceOutputs->haveForceWithVirial())
340 auto& forceWithVirial = forceOutputs->forceWithVirial();
344 /* Spread the mesh force on virtual sites to the other particles...
345 * This is parallellized. MPI communication is performed
346 * if the constructing atoms aren't local.
348 GMX_ASSERT(!stepWork.computeVirial || f.data() != forceWithVirial.force_.data(),
349 "We need separate force buffers for shift and virial forces when "
350 "computing the virial");
351 GMX_ASSERT(!stepWork.computeVirial
352 || forceOutputs->forceWithShiftForces().haveSpreadVsiteForces(),
353 "We should spread the force with shift forces separately when computing "
355 const gmx::VirtualSitesHandler::VirialHandling virialHandling =
356 (stepWork.computeVirial ? gmx::VirtualSitesHandler::VirialHandling::NonLinear
357 : gmx::VirtualSitesHandler::VirialHandling::None);
358 matrix virial = { { 0 } };
359 vsite->spreadForces(x, forceWithVirial.force_, virialHandling, {}, virial, nrnb, box, wcycle);
360 forceWithVirial.addVirialContribution(virial);
363 if (stepWork.computeVirial)
365 /* Now add the forces, this is local */
366 sum_forces(f, forceWithVirial.force_);
368 /* Add the direct virial contributions */
370 forceWithVirial.computeVirial_,
371 "forceWithVirial should request virial computation when we request the virial");
372 m_add(vir_force, forceWithVirial.getVirial(), vir_force);
376 pr_rvecs(debug, 0, "vir_force", vir_force, DIM);
382 GMX_ASSERT(vsite == nullptr || forceOutputs->forceWithShiftForces().haveSpreadVsiteForces(),
383 "We should have spread the vsite forces (earlier)");
386 if (fr->print_force >= 0)
388 print_large_forces(stderr, mdatoms, cr, step, fr->print_force, x, f);
392 static void do_nb_verlet(t_forcerec* fr,
393 const interaction_const_t* ic,
394 gmx_enerdata_t* enerd,
395 const StepWorkload& stepWork,
396 const InteractionLocality ilocality,
400 gmx_wallcycle_t wcycle)
402 if (!stepWork.computeNonbondedForces)
404 /* skip non-bonded calculation */
408 nonbonded_verlet_t* nbv = fr->nbv.get();
410 /* GPU kernel launch overhead is already timed separately */
411 if (fr->cutoff_scheme != ecutsVERLET)
413 gmx_incons("Invalid cut-off scheme passed!");
418 /* When dynamic pair-list pruning is requested, we need to prune
419 * at nstlistPrune steps.
421 if (nbv->isDynamicPruningStepCpu(step))
423 /* Prune the pair-list beyond fr->ic->rlistPrune using
424 * the current coordinates of the atoms.
426 wallcycle_sub_start(wcycle, ewcsNONBONDED_PRUNING);
427 nbv->dispatchPruneKernelCpu(ilocality, fr->shift_vec);
428 wallcycle_sub_stop(wcycle, ewcsNONBONDED_PRUNING);
432 nbv->dispatchNonbondedKernel(ilocality, *ic, stepWork, clearF, *fr, enerd, nrnb);
435 static inline void clearRVecs(ArrayRef<RVec> v, const bool useOpenmpThreading)
437 int nth = gmx_omp_nthreads_get_simple_rvec_task(emntDefault, v.ssize());
439 /* Note that we would like to avoid this conditional by putting it
440 * into the omp pragma instead, but then we still take the full
441 * omp parallel for overhead (at least with gcc5).
443 if (!useOpenmpThreading || nth == 1)
452 #pragma omp parallel for num_threads(nth) schedule(static)
453 for (gmx::index i = 0; i < v.ssize(); i++)
460 /*! \brief Return an estimate of the average kinetic energy or 0 when unreliable
462 * \param groupOptions Group options, containing T-coupling options
464 static real averageKineticEnergyEstimate(const t_grpopts& groupOptions)
466 real nrdfCoupled = 0;
467 real nrdfUncoupled = 0;
468 real kineticEnergy = 0;
469 for (int g = 0; g < groupOptions.ngtc; g++)
471 if (groupOptions.tau_t[g] >= 0)
473 nrdfCoupled += groupOptions.nrdf[g];
474 kineticEnergy += groupOptions.nrdf[g] * 0.5 * groupOptions.ref_t[g] * BOLTZ;
478 nrdfUncoupled += groupOptions.nrdf[g];
482 /* This conditional with > also catches nrdf=0 */
483 if (nrdfCoupled > nrdfUncoupled)
485 return kineticEnergy * (nrdfCoupled + nrdfUncoupled) / nrdfCoupled;
493 /*! \brief This routine checks that the potential energy is finite.
495 * Always checks that the potential energy is finite. If step equals
496 * inputrec.init_step also checks that the magnitude of the potential energy
497 * is reasonable. Terminates with a fatal error when a check fails.
498 * Note that passing this check does not guarantee finite forces,
499 * since those use slightly different arithmetics. But in most cases
500 * there is just a narrow coordinate range where forces are not finite
501 * and energies are finite.
503 * \param[in] step The step number, used for checking and printing
504 * \param[in] enerd The energy data; the non-bonded group energies need to be added to
505 * enerd.term[F_EPOT] before calling this routine \param[in] inputrec The input record
507 static void checkPotentialEnergyValidity(int64_t step, const gmx_enerdata_t& enerd, const t_inputrec& inputrec)
509 /* Threshold valid for comparing absolute potential energy against
510 * the kinetic energy. Normally one should not consider absolute
511 * potential energy values, but with a factor of one million
512 * we should never get false positives.
514 constexpr real c_thresholdFactor = 1e6;
516 bool energyIsNotFinite = !std::isfinite(enerd.term[F_EPOT]);
517 real averageKineticEnergy = 0;
518 /* We only check for large potential energy at the initial step,
519 * because that is by far the most likely step for this too occur
520 * and because computing the average kinetic energy is not free.
521 * Note: nstcalcenergy >> 1 often does not allow to catch large energies
522 * before they become NaN.
524 if (step == inputrec.init_step && EI_DYNAMICS(inputrec.eI))
526 averageKineticEnergy = averageKineticEnergyEstimate(inputrec.opts);
529 if (energyIsNotFinite
530 || (averageKineticEnergy > 0 && enerd.term[F_EPOT] > c_thresholdFactor * averageKineticEnergy))
535 ": The total potential energy is %g, which is %s. The LJ and electrostatic "
536 "contributions to the energy are %g and %g, respectively. A %s potential energy "
537 "can be caused by overlapping interactions in bonded interactions or very large%s "
538 "coordinate values. Usually this is caused by a badly- or non-equilibrated initial "
539 "configuration, incorrect interactions or parameters in the topology.",
540 step, enerd.term[F_EPOT], energyIsNotFinite ? "not finite" : "extremely high",
541 enerd.term[F_LJ], enerd.term[F_COUL_SR],
542 energyIsNotFinite ? "non-finite" : "very high", energyIsNotFinite ? " or Nan" : "");
546 /*! \brief Return true if there are special forces computed this step.
548 * The conditionals exactly correspond to those in computeSpecialForces().
550 static bool haveSpecialForces(const t_inputrec& inputrec,
551 const gmx::ForceProviders& forceProviders,
552 const pull_t* pull_work,
553 const bool computeForces,
557 return ((computeForces && forceProviders.hasForceProvider()) || // forceProviders
558 (inputrec.bPull && pull_have_potential(pull_work)) || // pull
559 inputrec.bRot || // enforced rotation
560 (ed != nullptr) || // flooding
561 (inputrec.bIMD && computeForces)); // IMD
564 /*! \brief Compute forces and/or energies for special algorithms
566 * The intention is to collect all calls to algorithms that compute
567 * forces on local atoms only and that do not contribute to the local
568 * virial sum (but add their virial contribution separately).
569 * Eventually these should likely all become ForceProviders.
570 * Within this function the intention is to have algorithms that do
571 * global communication at the end, so global barriers within the MD loop
572 * are as close together as possible.
574 * \param[in] fplog The log file
575 * \param[in] cr The communication record
576 * \param[in] inputrec The input record
577 * \param[in] awh The Awh module (nullptr if none in use).
578 * \param[in] enforcedRotation Enforced rotation module.
579 * \param[in] imdSession The IMD session
580 * \param[in] pull_work The pull work structure.
581 * \param[in] step The current MD step
582 * \param[in] t The current time
583 * \param[in,out] wcycle Wallcycle accounting struct
584 * \param[in,out] forceProviders Pointer to a list of force providers
585 * \param[in] box The unit cell
586 * \param[in] x The coordinates
587 * \param[in] mdatoms Per atom properties
588 * \param[in] lambda Array of free-energy lambda values
589 * \param[in] stepWork Step schedule flags
590 * \param[in,out] forceWithVirial Force and virial buffers
591 * \param[in,out] enerd Energy buffer
592 * \param[in,out] ed Essential dynamics pointer
593 * \param[in] didNeighborSearch Tells if we did neighbor searching this step, used for ED sampling
595 * \todo Remove didNeighborSearch, which is used incorrectly.
596 * \todo Convert all other algorithms called here to ForceProviders.
598 static void computeSpecialForces(FILE* fplog,
600 const t_inputrec* inputrec,
602 gmx_enfrot* enforcedRotation,
603 gmx::ImdSession* imdSession,
607 gmx_wallcycle_t wcycle,
608 gmx::ForceProviders* forceProviders,
610 gmx::ArrayRef<const gmx::RVec> x,
611 const t_mdatoms* mdatoms,
613 const StepWorkload& stepWork,
614 gmx::ForceWithVirial* forceWithVirial,
615 gmx_enerdata_t* enerd,
617 bool didNeighborSearch)
619 /* NOTE: Currently all ForceProviders only provide forces.
620 * When they also provide energies, remove this conditional.
622 if (stepWork.computeForces)
624 gmx::ForceProviderInput forceProviderInput(x, *mdatoms, t, box, *cr);
625 gmx::ForceProviderOutput forceProviderOutput(forceWithVirial, enerd);
627 /* Collect forces from modules */
628 forceProviders->calculateForces(forceProviderInput, &forceProviderOutput);
631 if (inputrec->bPull && pull_have_potential(pull_work))
633 pull_potential_wrapper(cr, inputrec, box, x, forceWithVirial, mdatoms, enerd, pull_work,
638 enerd->term[F_COM_PULL] += awh->applyBiasForcesAndUpdateBias(
639 inputrec->pbcType, mdatoms->massT, box, forceWithVirial, t, step, wcycle, fplog);
643 rvec* f = as_rvec_array(forceWithVirial->force_.data());
645 /* Add the forces from enforced rotation potentials (if any) */
648 wallcycle_start(wcycle, ewcROTadd);
649 enerd->term[F_COM_PULL] += add_rot_forces(enforcedRotation, f, cr, step, t);
650 wallcycle_stop(wcycle, ewcROTadd);
655 /* Note that since init_edsam() is called after the initialization
656 * of forcerec, edsam doesn't request the noVirSum force buffer.
657 * Thus if no other algorithm (e.g. PME) requires it, the forces
658 * here will contribute to the virial.
660 do_flood(cr, inputrec, as_rvec_array(x.data()), f, ed, box, step, didNeighborSearch);
663 /* Add forces from interactive molecular dynamics (IMD), if any */
664 if (inputrec->bIMD && stepWork.computeForces)
666 imdSession->applyForces(f);
670 /*! \brief Launch the prepare_step and spread stages of PME GPU.
672 * \param[in] pmedata The PME structure
673 * \param[in] box The box matrix
674 * \param[in] stepWork Step schedule flags
675 * \param[in] xReadyOnDevice Event synchronizer indicating that the coordinates are ready in
676 * the device memory. \param[in] wcycle The wallcycle structure
678 static inline void launchPmeGpuSpread(gmx_pme_t* pmedata,
680 const StepWorkload& stepWork,
681 GpuEventSynchronizer* xReadyOnDevice,
682 gmx_wallcycle_t wcycle)
684 pme_gpu_prepare_computation(pmedata, box, wcycle, stepWork);
685 pme_gpu_launch_spread(pmedata, xReadyOnDevice, wcycle);
688 /*! \brief Launch the FFT and gather stages of PME GPU
690 * This function only implements setting the output forces (no accumulation).
692 * \param[in] pmedata The PME structure
693 * \param[in] wcycle The wallcycle structure
694 * \param[in] stepWork Step schedule flags
696 static void launchPmeGpuFftAndGather(gmx_pme_t* pmedata, gmx_wallcycle_t wcycle, const gmx::StepWorkload& stepWork)
698 pme_gpu_launch_complex_transforms(pmedata, wcycle, stepWork);
699 pme_gpu_launch_gather(pmedata, wcycle);
703 * Polling wait for either of the PME or nonbonded GPU tasks.
705 * Instead of a static order in waiting for GPU tasks, this function
706 * polls checking which of the two tasks completes first, and does the
707 * associated force buffer reduction overlapped with the other task.
708 * By doing that, unlike static scheduling order, it can always overlap
709 * one of the reductions, regardless of the GPU task completion order.
711 * \param[in] nbv Nonbonded verlet structure
712 * \param[in,out] pmedata PME module data
713 * \param[in,out] forceOutputs Output buffer for the forces and virial
714 * \param[in,out] enerd Energy data structure results are reduced into
715 * \param[in] stepWork Step schedule flags
716 * \param[in] wcycle The wallcycle structure
718 static void alternatePmeNbGpuWaitReduce(nonbonded_verlet_t* nbv,
720 gmx::ForceOutputs* forceOutputs,
721 gmx_enerdata_t* enerd,
722 const StepWorkload& stepWork,
723 gmx_wallcycle_t wcycle)
725 bool isPmeGpuDone = false;
726 bool isNbGpuDone = false;
729 gmx::ForceWithShiftForces& forceWithShiftForces = forceOutputs->forceWithShiftForces();
730 gmx::ForceWithVirial& forceWithVirial = forceOutputs->forceWithVirial();
732 gmx::ArrayRef<const gmx::RVec> pmeGpuForces;
734 while (!isPmeGpuDone || !isNbGpuDone)
738 GpuTaskCompletion completionType =
739 (isNbGpuDone) ? GpuTaskCompletion::Wait : GpuTaskCompletion::Check;
740 isPmeGpuDone = pme_gpu_try_finish_task(pmedata, stepWork, wcycle, &forceWithVirial,
741 enerd, completionType);
746 GpuTaskCompletion completionType =
747 (isPmeGpuDone) ? GpuTaskCompletion::Wait : GpuTaskCompletion::Check;
748 isNbGpuDone = Nbnxm::gpu_try_finish_task(
749 nbv->gpu_nbv, stepWork, AtomLocality::Local, enerd->grpp.ener[egLJSR].data(),
750 enerd->grpp.ener[egCOULSR].data(), forceWithShiftForces.shiftForces(),
751 completionType, wcycle);
755 nbv->atomdata_add_nbat_f_to_f(AtomLocality::Local, forceWithShiftForces.force());
761 /*! \brief Set up the different force buffers; also does clearing.
763 * \param[in] forceHelperBuffers Helper force buffers
764 * \param[in] pull_work The pull work object.
765 * \param[in] inputrec input record
766 * \param[in] force force array
767 * \param[in] stepWork Step schedule flags
768 * \param[out] wcycle wallcycle recording structure
770 * \returns Cleared force output structure
772 static ForceOutputs setupForceOutputs(ForceHelperBuffers* forceHelperBuffers,
774 const t_inputrec& inputrec,
775 gmx::ArrayRefWithPadding<gmx::RVec> force,
776 const StepWorkload& stepWork,
777 gmx_wallcycle_t wcycle)
779 wallcycle_sub_start(wcycle, ewcsCLEAR_FORCE_BUFFER);
781 /* NOTE: We assume fr->shiftForces is all zeros here */
782 gmx::ForceWithShiftForces forceWithShiftForces(force, stepWork.computeVirial,
783 forceHelperBuffers->shiftForces());
785 if (stepWork.computeForces)
787 /* Clear the short- and long-range forces */
788 clearRVecs(forceWithShiftForces.force(), true);
790 /* Clear the shift forces */
791 clearRVecs(forceWithShiftForces.shiftForces(), false);
794 /* If we need to compute the virial, we might need a separate
795 * force buffer for algorithms for which the virial is calculated
796 * directly, such as PME. Otherwise, forceWithVirial uses the
797 * the same force (f in legacy calls) buffer as other algorithms.
799 const bool useSeparateForceWithVirialBuffer =
800 (stepWork.computeForces
801 && (stepWork.computeVirial && forceHelperBuffers->haveDirectVirialContributions()));
802 /* forceWithVirial uses the local atom range only */
803 gmx::ForceWithVirial forceWithVirial(
804 useSeparateForceWithVirialBuffer ? forceHelperBuffers->forceBufferForDirectVirialContributions()
805 : force.unpaddedArrayRef(),
806 stepWork.computeVirial);
808 if (useSeparateForceWithVirialBuffer)
810 /* TODO: update comment
811 * We only compute forces on local atoms. Note that vsites can
812 * spread to non-local atoms, but that part of the buffer is
813 * cleared separately in the vsite spreading code.
815 clearRVecs(forceWithVirial.force_, true);
818 if (inputrec.bPull && pull_have_constraint(pull_work))
820 clear_pull_forces(pull_work);
823 wallcycle_sub_stop(wcycle, ewcsCLEAR_FORCE_BUFFER);
825 return ForceOutputs(forceWithShiftForces, forceHelperBuffers->haveDirectVirialContributions(),
830 /*! \brief Set up flags that have the lifetime of the domain indicating what type of work is there to compute.
832 static DomainLifetimeWorkload setupDomainLifetimeWorkload(const t_inputrec& inputrec,
833 const t_forcerec& fr,
834 const pull_t* pull_work,
836 const t_mdatoms& mdatoms,
837 const SimulationWorkload& simulationWork,
838 const StepWorkload& stepWork)
840 DomainLifetimeWorkload domainWork;
841 // Note that haveSpecialForces is constant over the whole run
842 domainWork.haveSpecialForces =
843 haveSpecialForces(inputrec, *fr.forceProviders, pull_work, stepWork.computeForces, ed);
844 domainWork.haveCpuBondedWork = fr.listedForces->haveCpuBondeds();
845 domainWork.haveGpuBondedWork = ((fr.gpuBonded != nullptr) && fr.gpuBonded->haveInteractions());
846 domainWork.haveRestraintsWork = fr.listedForces->haveRestraints();
847 domainWork.haveCpuListedForceWork = fr.listedForces->haveCpuListedForces();
848 // Note that haveFreeEnergyWork is constant over the whole run
849 domainWork.haveFreeEnergyWork = (fr.efep != efepNO && mdatoms.nPerturbed != 0);
850 // We assume we have local force work if there are CPU
851 // force tasks including PME or nonbondeds.
852 domainWork.haveCpuLocalForceWork =
853 domainWork.haveSpecialForces || domainWork.haveCpuListedForceWork
854 || domainWork.haveFreeEnergyWork || simulationWork.useCpuNonbonded || simulationWork.useCpuPme
855 || simulationWork.haveEwaldSurfaceContribution || inputrec.nwall > 0;
860 /*! \brief Set up force flag stuct from the force bitmask.
862 * \param[in] legacyFlags Force bitmask flags used to construct the new flags
863 * \param[in] isNonbondedOn Global override, if false forces to turn off all nonbonded calculation.
864 * \param[in] simulationWork Simulation workload description.
865 * \param[in] rankHasPmeDuty If this rank computes PME.
867 * \returns New Stepworkload description.
869 static StepWorkload setupStepWorkload(const int legacyFlags,
870 const bool isNonbondedOn,
871 const SimulationWorkload& simulationWork,
872 const bool rankHasPmeDuty)
875 flags.stateChanged = ((legacyFlags & GMX_FORCE_STATECHANGED) != 0);
876 flags.haveDynamicBox = ((legacyFlags & GMX_FORCE_DYNAMICBOX) != 0);
877 flags.doNeighborSearch = ((legacyFlags & GMX_FORCE_NS) != 0);
878 flags.computeVirial = ((legacyFlags & GMX_FORCE_VIRIAL) != 0);
879 flags.computeEnergy = ((legacyFlags & GMX_FORCE_ENERGY) != 0);
880 flags.computeForces = ((legacyFlags & GMX_FORCE_FORCES) != 0);
881 flags.computeListedForces = ((legacyFlags & GMX_FORCE_LISTED) != 0);
882 flags.computeNonbondedForces = ((legacyFlags & GMX_FORCE_NONBONDED) != 0) && isNonbondedOn;
883 flags.computeDhdl = ((legacyFlags & GMX_FORCE_DHDL) != 0);
885 if (simulationWork.useGpuBufferOps)
887 GMX_ASSERT(simulationWork.useGpuNonbonded,
888 "Can only offload buffer ops if nonbonded computation is also offloaded");
890 flags.useGpuXBufferOps = simulationWork.useGpuBufferOps;
891 // on virial steps the CPU reduction path is taken
892 flags.useGpuFBufferOps = simulationWork.useGpuBufferOps && !flags.computeVirial;
893 flags.useGpuPmeFReduction = flags.useGpuFBufferOps
894 && (simulationWork.useGpuPme
895 && (rankHasPmeDuty || simulationWork.useGpuPmePpCommunication));
901 /* \brief Launch end-of-step GPU tasks: buffer clearing and rolling pruning.
903 * TODO: eliminate \p useGpuPmeOnThisRank when this is
904 * incorporated in DomainLifetimeWorkload.
906 static void launchGpuEndOfStepTasks(nonbonded_verlet_t* nbv,
907 gmx::GpuBonded* gpuBonded,
909 gmx_enerdata_t* enerd,
910 const gmx::MdrunScheduleWorkload& runScheduleWork,
911 bool useGpuPmeOnThisRank,
913 gmx_wallcycle_t wcycle)
915 if (runScheduleWork.simulationWork.useGpuNonbonded)
917 /* Launch pruning before buffer clearing because the API overhead of the
918 * clear kernel launches can leave the GPU idle while it could be running
921 if (nbv->isDynamicPruningStepGpu(step))
923 nbv->dispatchPruneKernelGpu(step);
926 /* now clear the GPU outputs while we finish the step on the CPU */
927 wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
928 wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
929 Nbnxm::gpu_clear_outputs(nbv->gpu_nbv, runScheduleWork.stepWork.computeVirial);
930 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
931 wallcycle_stop(wcycle, ewcLAUNCH_GPU);
934 if (useGpuPmeOnThisRank)
936 pme_gpu_reinit_computation(pmedata, wcycle);
939 if (runScheduleWork.domainWork.haveGpuBondedWork && runScheduleWork.stepWork.computeEnergy)
941 // in principle this should be included in the DD balancing region,
942 // but generally it is infrequent so we'll omit it for the sake of
944 gpuBonded->waitAccumulateEnergyTerms(enerd);
946 gpuBonded->clearEnergies();
950 //! \brief Data structure to hold dipole-related data and staging arrays
953 //! Dipole staging for fast summing over MPI
954 gmx::DVec muStaging[2] = { { 0.0, 0.0, 0.0 } };
955 //! Dipole staging for states A and B (index 0 and 1 resp.)
956 gmx::RVec muStateAB[2] = { { 0.0_real, 0.0_real, 0.0_real } };
960 static void reduceAndUpdateMuTot(DipoleData* dipoleData,
962 const bool haveFreeEnergy,
963 gmx::ArrayRef<const real> lambda,
965 const DDBalanceRegionHandler& ddBalanceRegionHandler)
969 gmx_sumd(2 * DIM, dipoleData->muStaging[0], cr);
970 ddBalanceRegionHandler.reopenRegionCpu();
972 for (int i = 0; i < 2; i++)
974 for (int j = 0; j < DIM; j++)
976 dipoleData->muStateAB[i][j] = dipoleData->muStaging[i][j];
982 copy_rvec(dipoleData->muStateAB[0], muTotal);
986 for (int j = 0; j < DIM; j++)
988 muTotal[j] = (1.0 - lambda[efptCOUL]) * dipoleData->muStateAB[0][j]
989 + lambda[efptCOUL] * dipoleData->muStateAB[1][j];
994 void do_force(FILE* fplog,
996 const gmx_multisim_t* ms,
997 const t_inputrec* inputrec,
999 gmx_enfrot* enforcedRotation,
1000 gmx::ImdSession* imdSession,
1004 gmx_wallcycle_t wcycle,
1005 const gmx_localtop_t* top,
1007 gmx::ArrayRefWithPadding<gmx::RVec> x,
1009 gmx::ArrayRefWithPadding<gmx::RVec> force,
1011 const t_mdatoms* mdatoms,
1012 gmx_enerdata_t* enerd,
1013 gmx::ArrayRef<real> lambda,
1015 gmx::MdrunScheduleWorkload* runScheduleWork,
1016 gmx::VirtualSitesHandler* vsite,
1021 const DDBalanceRegionHandler& ddBalanceRegionHandler)
1023 GMX_ASSERT(force.unpaddedArrayRef().ssize() >= fr->natoms_force_constr,
1024 "The size of the force buffer should be at least the number of atoms to compute "
1027 nonbonded_verlet_t* nbv = fr->nbv.get();
1028 interaction_const_t* ic = fr->ic;
1029 gmx::StatePropagatorDataGpu* stateGpu = fr->stateGpu;
1031 const SimulationWorkload& simulationWork = runScheduleWork->simulationWork;
1034 runScheduleWork->stepWork = setupStepWorkload(legacyFlags, fr->bNonbonded, simulationWork,
1035 thisRankHasDuty(cr, DUTY_PME));
1036 const StepWorkload& stepWork = runScheduleWork->stepWork;
1039 const bool useGpuPmeOnThisRank = simulationWork.useGpuPme && thisRankHasDuty(cr, DUTY_PME);
1041 /* At a search step we need to start the first balancing region
1042 * somewhere early inside the step after communication during domain
1043 * decomposition (and not during the previous step as usual).
1045 if (stepWork.doNeighborSearch)
1047 ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::yes);
1050 clear_mat(vir_force);
1052 if (fr->pbcType != PbcType::No)
1054 /* Compute shift vectors every step,
1055 * because of pressure coupling or box deformation!
1057 if (stepWork.haveDynamicBox && stepWork.stateChanged)
1059 calc_shifts(box, fr->shift_vec);
1062 const bool fillGrid = (stepWork.doNeighborSearch && stepWork.stateChanged);
1063 const bool calcCGCM = (fillGrid && !DOMAINDECOMP(cr));
1066 put_atoms_in_box_omp(fr->pbcType, box, x.unpaddedArrayRef().subArray(0, mdatoms->homenr),
1067 gmx_omp_nthreads_get(emntDefault));
1068 inc_nrnb(nrnb, eNR_SHIFTX, mdatoms->homenr);
1072 nbnxn_atomdata_copy_shiftvec(stepWork.haveDynamicBox, fr->shift_vec, nbv->nbat.get());
1074 const bool pmeSendCoordinatesFromGpu =
1075 GMX_MPI && simulationWork.useGpuPmePpCommunication && !(stepWork.doNeighborSearch);
1076 const bool reinitGpuPmePpComms =
1077 GMX_MPI && simulationWork.useGpuPmePpCommunication && (stepWork.doNeighborSearch);
1079 const auto localXReadyOnDevice = (useGpuPmeOnThisRank || simulationWork.useGpuBufferOps)
1080 ? stateGpu->getCoordinatesReadyOnDeviceEvent(
1081 AtomLocality::Local, simulationWork, stepWork)
1084 // If coordinates are to be sent to PME task from CPU memory, perform that send here.
1085 // Otherwise the send will occur after H2D coordinate transfer.
1086 if (GMX_MPI && !thisRankHasDuty(cr, DUTY_PME) && !pmeSendCoordinatesFromGpu)
1088 /* Send particle coordinates to the pme nodes */
1089 if (!stepWork.doNeighborSearch && simulationWork.useGpuUpdate)
1091 GMX_RELEASE_ASSERT(false,
1092 "GPU update and separate PME ranks are only supported with GPU "
1093 "direct communication!");
1094 // TODO: when this code-path becomes supported add:
1095 // stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
1098 gmx_pme_send_coordinates(fr, cr, box, as_rvec_array(x.unpaddedArrayRef().data()), lambda[efptCOUL],
1099 lambda[efptVDW], (stepWork.computeVirial || stepWork.computeEnergy),
1100 step, simulationWork.useGpuPmePpCommunication, reinitGpuPmePpComms,
1101 pmeSendCoordinatesFromGpu, localXReadyOnDevice, wcycle);
1104 // Coordinates on the device are needed if PME or BufferOps are offloaded.
1105 // The local coordinates can be copied right away.
1106 // NOTE: Consider moving this copy to right after they are updated and constrained,
1107 // if the later is not offloaded.
1108 if (useGpuPmeOnThisRank || stepWork.useGpuXBufferOps)
1110 if (stepWork.doNeighborSearch)
1112 // TODO refactor this to do_md, after partitioning.
1113 stateGpu->reinit(mdatoms->homenr,
1114 cr->dd != nullptr ? dd_numAtomsZones(*cr->dd) : mdatoms->homenr);
1115 if (useGpuPmeOnThisRank)
1117 // TODO: This should be moved into PME setup function ( pme_gpu_prepare_computation(...) )
1118 pme_gpu_set_device_x(fr->pmedata, stateGpu->getCoordinates());
1121 // We need to copy coordinates when:
1122 // 1. Update is not offloaded
1123 // 2. The buffers were reinitialized on search step
1124 if (!simulationWork.useGpuUpdate || stepWork.doNeighborSearch)
1126 GMX_ASSERT(stateGpu != nullptr, "stateGpu should not be null");
1127 stateGpu->copyCoordinatesToGpu(x.unpaddedArrayRef(), AtomLocality::Local);
1131 // TODO Update this comment when introducing SimulationWorkload
1133 // The conditions for gpuHaloExchange e.g. using GPU buffer
1134 // operations were checked before construction, so here we can
1135 // just use it and assert upon any conditions.
1136 const bool ddUsesGpuDirectCommunication =
1137 ((cr->dd != nullptr) && (!cr->dd->gpuHaloExchange.empty()));
1138 GMX_ASSERT(!ddUsesGpuDirectCommunication || stepWork.useGpuXBufferOps,
1139 "Must use coordinate buffer ops with GPU halo exchange");
1140 const bool useGpuForcesHaloExchange = ddUsesGpuDirectCommunication && stepWork.useGpuFBufferOps;
1142 // Copy coordinate from the GPU if update is on the GPU and there
1143 // are forces to be computed on the CPU, or for the computation of
1144 // virial, or if host-side data will be transferred from this task
1145 // to a remote task for halo exchange or PME-PP communication. At
1146 // search steps the current coordinates are already on the host,
1147 // hence copy is not needed.
1148 const bool haveHostPmePpComms =
1149 !thisRankHasDuty(cr, DUTY_PME) && !simulationWork.useGpuPmePpCommunication;
1150 const bool haveHostHaloExchangeComms = havePPDomainDecomposition(cr) && !ddUsesGpuDirectCommunication;
1152 bool gmx_used_in_debug haveCopiedXFromGpu = false;
1153 if (simulationWork.useGpuUpdate && !stepWork.doNeighborSearch
1154 && (runScheduleWork->domainWork.haveCpuLocalForceWork || stepWork.computeVirial
1155 || haveHostPmePpComms || haveHostHaloExchangeComms))
1157 GMX_ASSERT(stateGpu != nullptr, "stateGpu should not be null");
1158 stateGpu->copyCoordinatesFromGpu(x.unpaddedArrayRef(), AtomLocality::Local);
1159 haveCopiedXFromGpu = true;
1162 // If coordinates are to be sent to PME task from GPU memory, perform that send here.
1163 // Otherwise the send will occur before the H2D coordinate transfer.
1164 if (!thisRankHasDuty(cr, DUTY_PME) && pmeSendCoordinatesFromGpu)
1166 /* Send particle coordinates to the pme nodes */
1167 gmx_pme_send_coordinates(fr, cr, box, as_rvec_array(x.unpaddedArrayRef().data()), lambda[efptCOUL],
1168 lambda[efptVDW], (stepWork.computeVirial || stepWork.computeEnergy),
1169 step, simulationWork.useGpuPmePpCommunication, reinitGpuPmePpComms,
1170 pmeSendCoordinatesFromGpu, localXReadyOnDevice, wcycle);
1173 if (useGpuPmeOnThisRank)
1175 launchPmeGpuSpread(fr->pmedata, box, stepWork, localXReadyOnDevice, wcycle);
1178 /* do gridding for pair search */
1179 if (stepWork.doNeighborSearch)
1181 if (fr->wholeMoleculeTransform && stepWork.stateChanged)
1183 fr->wholeMoleculeTransform->updateForAtomPbcJumps(x.unpaddedArrayRef(), box);
1187 // - vzero is constant, do we need to pass it?
1188 // - box_diag should be passed directly to nbnxn_put_on_grid
1194 box_diag[XX] = box[XX][XX];
1195 box_diag[YY] = box[YY][YY];
1196 box_diag[ZZ] = box[ZZ][ZZ];
1198 wallcycle_start(wcycle, ewcNS);
1199 if (!DOMAINDECOMP(cr))
1201 wallcycle_sub_start(wcycle, ewcsNBS_GRID_LOCAL);
1202 nbnxn_put_on_grid(nbv, box, 0, vzero, box_diag, nullptr, { 0, mdatoms->homenr }, -1,
1203 fr->cginfo, x.unpaddedArrayRef(), 0, nullptr);
1204 wallcycle_sub_stop(wcycle, ewcsNBS_GRID_LOCAL);
1208 wallcycle_sub_start(wcycle, ewcsNBS_GRID_NONLOCAL);
1209 nbnxn_put_on_grid_nonlocal(nbv, domdec_zones(cr->dd), fr->cginfo, x.unpaddedArrayRef());
1210 wallcycle_sub_stop(wcycle, ewcsNBS_GRID_NONLOCAL);
1213 nbv->setAtomProperties(gmx::constArrayRefFromArray(mdatoms->typeA, mdatoms->nr),
1214 gmx::constArrayRefFromArray(mdatoms->chargeA, mdatoms->nr), fr->cginfo);
1216 wallcycle_stop(wcycle, ewcNS);
1218 /* initialize the GPU nbnxm atom data and bonded data structures */
1219 if (simulationWork.useGpuNonbonded)
1221 wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1223 wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1224 Nbnxm::gpu_init_atomdata(nbv->gpu_nbv, nbv->nbat.get());
1225 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1229 /* Now we put all atoms on the grid, we can assign bonded
1230 * interactions to the GPU, where the grid order is
1231 * needed. Also the xq, f and fshift device buffers have
1232 * been reallocated if needed, so the bonded code can
1233 * learn about them. */
1234 // TODO the xq, f, and fshift buffers are now shared
1235 // resources, so they should be maintained by a
1236 // higher-level object than the nb module.
1237 fr->gpuBonded->updateInteractionListsAndDeviceBuffers(
1238 nbv->getGridIndices(), top->idef, Nbnxm::gpu_get_xq(nbv->gpu_nbv),
1239 Nbnxm::gpu_get_f(nbv->gpu_nbv), Nbnxm::gpu_get_fshift(nbv->gpu_nbv));
1241 wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1244 // Need to run after the GPU-offload bonded interaction lists
1245 // are set up to be able to determine whether there is bonded work.
1246 runScheduleWork->domainWork = setupDomainLifetimeWorkload(
1247 *inputrec, *fr, pull_work, ed, *mdatoms, simulationWork, stepWork);
1249 wallcycle_start_nocount(wcycle, ewcNS);
1250 wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_LOCAL);
1251 /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
1252 nbv->constructPairlist(InteractionLocality::Local, top->excls, step, nrnb);
1254 nbv->setupGpuShortRangeWork(fr->gpuBonded, InteractionLocality::Local);
1256 wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_LOCAL);
1257 wallcycle_stop(wcycle, ewcNS);
1259 if (stepWork.useGpuXBufferOps)
1261 nbv->atomdata_init_copy_x_to_nbat_x_gpu();
1263 // For force buffer ops, we use the below conditon rather than
1264 // useGpuFBufferOps to ensure that init is performed even if this
1265 // NS step is also a virial step (on which f buf ops are deactivated).
1266 if (simulationWork.useGpuBufferOps && simulationWork.useGpuNonbonded && (GMX_GPU == GMX_GPU_CUDA))
1268 GMX_ASSERT(stateGpu, "stateGpu should be valid when buffer ops are offloaded");
1269 nbv->atomdata_init_add_nbat_f_to_f_gpu(stateGpu->fReducedOnDevice());
1272 else if (!EI_TPI(inputrec->eI))
1274 if (stepWork.useGpuXBufferOps)
1276 GMX_ASSERT(stateGpu, "stateGpu should be valid when buffer ops are offloaded");
1277 nbv->convertCoordinatesGpu(AtomLocality::Local, false, stateGpu->getCoordinates(),
1278 localXReadyOnDevice);
1282 if (simulationWork.useGpuUpdate)
1284 GMX_ASSERT(stateGpu, "need a valid stateGpu object");
1285 GMX_ASSERT(haveCopiedXFromGpu,
1286 "a wait should only be triggered if copy has been scheduled");
1287 stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
1289 nbv->convertCoordinates(AtomLocality::Local, false, x.unpaddedArrayRef());
1293 const gmx::DomainLifetimeWorkload& domainWork = runScheduleWork->domainWork;
1295 if (simulationWork.useGpuNonbonded)
1297 ddBalanceRegionHandler.openBeforeForceComputationGpu();
1299 wallcycle_start(wcycle, ewcLAUNCH_GPU);
1301 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1302 Nbnxm::gpu_upload_shiftvec(nbv->gpu_nbv, nbv->nbat.get());
1303 if (stepWork.doNeighborSearch || !stepWork.useGpuXBufferOps)
1305 Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(), AtomLocality::Local);
1307 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1308 // with X buffer ops offloaded to the GPU on all but the search steps
1310 // bonded work not split into separate local and non-local, so with DD
1311 // we can only launch the kernel after non-local coordinates have been received.
1312 if (domainWork.haveGpuBondedWork && !havePPDomainDecomposition(cr))
1314 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_BONDED);
1315 fr->gpuBonded->setPbcAndlaunchKernel(fr->pbcType, box, fr->bMolPBC, stepWork);
1316 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1319 /* launch local nonbonded work on GPU */
1320 wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1321 do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::Local, enbvClearFNo, step, nrnb, wcycle);
1322 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1323 wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1326 if (useGpuPmeOnThisRank)
1328 // In PME GPU and mixed mode we launch FFT / gather after the
1329 // X copy/transform to allow overlap as well as after the GPU NB
1330 // launch to avoid FFT launch overhead hijacking the CPU and delaying
1331 // the nonbonded kernel.
1332 launchPmeGpuFftAndGather(fr->pmedata, wcycle, stepWork);
1335 /* Communicate coordinates and sum dipole if necessary +
1336 do non-local pair search */
1337 if (havePPDomainDecomposition(cr))
1339 if (stepWork.doNeighborSearch)
1341 // TODO: fuse this branch with the above large stepWork.doNeighborSearch block
1342 wallcycle_start_nocount(wcycle, ewcNS);
1343 wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1344 /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
1345 nbv->constructPairlist(InteractionLocality::NonLocal, top->excls, step, nrnb);
1347 nbv->setupGpuShortRangeWork(fr->gpuBonded, InteractionLocality::NonLocal);
1348 wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1349 wallcycle_stop(wcycle, ewcNS);
1350 // TODO refactor this GPU halo exchange re-initialisation
1351 // to location in do_md where GPU halo exchange is
1352 // constructed at partitioning, after above stateGpu
1353 // re-initialization has similarly been refactored
1354 if (ddUsesGpuDirectCommunication)
1356 reinitGpuHaloExchange(*cr, stateGpu->getCoordinates(), stateGpu->getForces());
1361 if (ddUsesGpuDirectCommunication)
1363 // The following must be called after local setCoordinates (which records an event
1364 // when the coordinate data has been copied to the device).
1365 communicateGpuHaloCoordinates(*cr, box, localXReadyOnDevice);
1367 if (domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork)
1369 // non-local part of coordinate buffer must be copied back to host for CPU work
1370 stateGpu->copyCoordinatesFromGpu(x.unpaddedArrayRef(), AtomLocality::NonLocal);
1375 // Note: GPU update + DD without direct communication is not supported,
1376 // a waitCoordinatesReadyOnHost() should be issued if it will be.
1377 GMX_ASSERT(!simulationWork.useGpuUpdate,
1378 "GPU update is not supported with CPU halo exchange");
1379 dd_move_x(cr->dd, box, x.unpaddedArrayRef(), wcycle);
1382 if (stepWork.useGpuXBufferOps)
1384 if (!useGpuPmeOnThisRank && !ddUsesGpuDirectCommunication)
1386 stateGpu->copyCoordinatesToGpu(x.unpaddedArrayRef(), AtomLocality::NonLocal);
1388 nbv->convertCoordinatesGpu(AtomLocality::NonLocal, false, stateGpu->getCoordinates(),
1389 stateGpu->getCoordinatesReadyOnDeviceEvent(
1390 AtomLocality::NonLocal, simulationWork, stepWork));
1394 nbv->convertCoordinates(AtomLocality::NonLocal, false, x.unpaddedArrayRef());
1398 if (simulationWork.useGpuNonbonded)
1400 wallcycle_start(wcycle, ewcLAUNCH_GPU);
1402 if (stepWork.doNeighborSearch || !stepWork.useGpuXBufferOps)
1404 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1405 Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(), AtomLocality::NonLocal);
1406 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1409 if (domainWork.haveGpuBondedWork)
1411 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_BONDED);
1412 fr->gpuBonded->setPbcAndlaunchKernel(fr->pbcType, box, fr->bMolPBC, stepWork);
1413 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1416 /* launch non-local nonbonded tasks on GPU */
1417 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1418 do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::NonLocal, enbvClearFNo, step,
1420 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1422 wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1426 if (simulationWork.useGpuNonbonded)
1428 /* launch D2H copy-back F */
1429 wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1430 wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1432 if (havePPDomainDecomposition(cr))
1434 Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat.get(), stepWork, AtomLocality::NonLocal);
1436 Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat.get(), stepWork, AtomLocality::Local);
1437 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1439 if (domainWork.haveGpuBondedWork && stepWork.computeEnergy)
1441 fr->gpuBonded->launchEnergyTransfer();
1443 wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1446 gmx::ArrayRef<const gmx::RVec> xWholeMolecules;
1447 if (fr->wholeMoleculeTransform)
1449 xWholeMolecules = fr->wholeMoleculeTransform->wholeMoleculeCoordinates(x.unpaddedArrayRef(), box);
1452 DipoleData dipoleData;
1454 if (simulationWork.computeMuTot)
1456 const int start = 0;
1458 /* Calculate total (local) dipole moment in a temporary common array.
1459 * This makes it possible to sum them over nodes faster.
1461 gmx::ArrayRef<const gmx::RVec> xRef =
1462 (xWholeMolecules.empty() ? x.unpaddedArrayRef() : xWholeMolecules);
1463 calc_mu(start, mdatoms->homenr, xRef, mdatoms->chargeA, mdatoms->chargeB,
1464 mdatoms->nChargePerturbed, dipoleData.muStaging[0], dipoleData.muStaging[1]);
1466 reduceAndUpdateMuTot(&dipoleData, cr, (fr->efep != efepNO), lambda, muTotal, ddBalanceRegionHandler);
1469 /* Reset energies */
1470 reset_enerdata(enerd);
1472 if (DOMAINDECOMP(cr) && !thisRankHasDuty(cr, DUTY_PME))
1474 wallcycle_start(wcycle, ewcPPDURINGPME);
1475 dd_force_flop_start(cr->dd, nrnb);
1478 // For the rest of the CPU tasks that depend on GPU-update produced coordinates,
1479 // this wait ensures that the D2H transfer is complete.
1480 if ((simulationWork.useGpuUpdate)
1481 && (runScheduleWork->domainWork.haveCpuLocalForceWork || stepWork.computeVirial))
1483 stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
1488 wallcycle_start(wcycle, ewcROT);
1489 do_rotation(cr, enforcedRotation, box, as_rvec_array(x.unpaddedArrayRef().data()), t, step,
1490 stepWork.doNeighborSearch);
1491 wallcycle_stop(wcycle, ewcROT);
1494 /* Start the force cycle counter.
1495 * Note that a different counter is used for dynamic load balancing.
1497 wallcycle_start(wcycle, ewcFORCE);
1499 // Set up and clear force outputs.
1500 // We use std::move to keep the compiler happy, it has no effect.
1501 ForceOutputs forceOut = setupForceOutputs(fr->forceHelperBuffers.get(), pull_work, *inputrec,
1502 std::move(force), stepWork, wcycle);
1504 /* We calculate the non-bonded forces, when done on the CPU, here.
1505 * We do this before calling do_force_lowlevel, because in that
1506 * function, the listed forces are calculated before PME, which
1507 * does communication. With this order, non-bonded and listed
1508 * force calculation imbalance can be balanced out by the domain
1509 * decomposition load balancing.
1512 const bool useOrEmulateGpuNb = simulationWork.useGpuNonbonded || fr->nbv->emulateGpu();
1514 if (!useOrEmulateGpuNb)
1516 do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::Local, enbvClearFYes, step, nrnb, wcycle);
1519 if (fr->efep != efepNO)
1521 /* Calculate the local and non-local free energy interactions here.
1522 * Happens here on the CPU both with and without GPU.
1524 nbv->dispatchFreeEnergyKernel(InteractionLocality::Local, fr,
1525 as_rvec_array(x.unpaddedArrayRef().data()),
1526 &forceOut.forceWithShiftForces(), *mdatoms, inputrec->fepvals,
1527 lambda.data(), enerd, stepWork, nrnb);
1529 if (havePPDomainDecomposition(cr))
1531 nbv->dispatchFreeEnergyKernel(InteractionLocality::NonLocal, fr,
1532 as_rvec_array(x.unpaddedArrayRef().data()),
1533 &forceOut.forceWithShiftForces(), *mdatoms,
1534 inputrec->fepvals, lambda.data(), enerd, stepWork, nrnb);
1538 if (!useOrEmulateGpuNb)
1540 if (havePPDomainDecomposition(cr))
1542 do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::NonLocal, enbvClearFNo, step,
1546 if (stepWork.computeForces)
1548 /* Add all the non-bonded force to the normal force array.
1549 * This can be split into a local and a non-local part when overlapping
1550 * communication with calculation with domain decomposition.
1552 wallcycle_stop(wcycle, ewcFORCE);
1553 nbv->atomdata_add_nbat_f_to_f(AtomLocality::All, forceOut.forceWithShiftForces().force());
1554 wallcycle_start_nocount(wcycle, ewcFORCE);
1557 /* If there are multiple fshift output buffers we need to reduce them */
1558 if (stepWork.computeVirial)
1560 /* This is not in a subcounter because it takes a
1561 negligible and constant-sized amount of time */
1562 nbnxn_atomdata_add_nbat_fshift_to_fshift(*nbv->nbat,
1563 forceOut.forceWithShiftForces().shiftForces());
1567 // TODO Force flags should include haveFreeEnergyWork for this domain
1568 if (ddUsesGpuDirectCommunication && (domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork))
1570 /* Wait for non-local coordinate data to be copied from device */
1571 stateGpu->waitCoordinatesReadyOnHost(AtomLocality::NonLocal);
1573 /* Compute the bonded and non-bonded energies and optionally forces */
1574 do_force_lowlevel(fr, inputrec, cr, ms, nrnb, wcycle, mdatoms, x, xWholeMolecules, hist,
1575 &forceOut, enerd, box, lambda.data(), as_rvec_array(dipoleData.muStateAB),
1576 stepWork, ddBalanceRegionHandler);
1578 wallcycle_stop(wcycle, ewcFORCE);
1580 computeSpecialForces(fplog, cr, inputrec, awh, enforcedRotation, imdSession, pull_work, step, t,
1581 wcycle, fr->forceProviders, box, x.unpaddedArrayRef(), mdatoms, lambda.data(),
1582 stepWork, &forceOut.forceWithVirial(), enerd, ed, stepWork.doNeighborSearch);
1585 // Will store the amount of cycles spent waiting for the GPU that
1586 // will be later used in the DLB accounting.
1587 float cycles_wait_gpu = 0;
1588 if (useOrEmulateGpuNb)
1590 auto& forceWithShiftForces = forceOut.forceWithShiftForces();
1592 /* wait for non-local forces (or calculate in emulation mode) */
1593 if (havePPDomainDecomposition(cr))
1595 if (simulationWork.useGpuNonbonded)
1597 cycles_wait_gpu += Nbnxm::gpu_wait_finish_task(
1598 nbv->gpu_nbv, stepWork, AtomLocality::NonLocal, enerd->grpp.ener[egLJSR].data(),
1599 enerd->grpp.ener[egCOULSR].data(), forceWithShiftForces.shiftForces(), wcycle);
1603 wallcycle_start_nocount(wcycle, ewcFORCE);
1604 do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::NonLocal, enbvClearFYes,
1605 step, nrnb, wcycle);
1606 wallcycle_stop(wcycle, ewcFORCE);
1609 if (stepWork.useGpuFBufferOps)
1611 gmx::FixedCapacityVector<GpuEventSynchronizer*, 1> dependencyList;
1613 // TODO: move this into DomainLifetimeWorkload, including the second part of the
1614 // condition The bonded and free energy CPU tasks can have non-local force
1615 // contributions which are a dependency for the GPU force reduction.
1616 bool haveNonLocalForceContribInCpuBuffer =
1617 domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork;
1619 if (haveNonLocalForceContribInCpuBuffer)
1621 stateGpu->copyForcesToGpu(forceOut.forceWithShiftForces().force(),
1622 AtomLocality::NonLocal);
1623 dependencyList.push_back(stateGpu->getForcesReadyOnDeviceEvent(
1624 AtomLocality::NonLocal, stepWork.useGpuFBufferOps));
1627 nbv->atomdata_add_nbat_f_to_f_gpu(AtomLocality::NonLocal, stateGpu->getForces(),
1628 pme_gpu_get_device_f(fr->pmedata), dependencyList,
1629 false, haveNonLocalForceContribInCpuBuffer);
1630 if (!useGpuForcesHaloExchange)
1632 // copy from GPU input for dd_move_f()
1633 stateGpu->copyForcesFromGpu(forceOut.forceWithShiftForces().force(),
1634 AtomLocality::NonLocal);
1639 nbv->atomdata_add_nbat_f_to_f(AtomLocality::NonLocal, forceWithShiftForces.force());
1643 if (fr->nbv->emulateGpu() && stepWork.computeVirial)
1645 nbnxn_atomdata_add_nbat_fshift_to_fshift(*nbv->nbat, forceWithShiftForces.shiftForces());
1650 if (havePPDomainDecomposition(cr))
1652 /* We are done with the CPU compute.
1653 * We will now communicate the non-local forces.
1654 * If we use a GPU this will overlap with GPU work, so in that case
1655 * we do not close the DD force balancing region here.
1657 ddBalanceRegionHandler.closeAfterForceComputationCpu();
1659 if (stepWork.computeForces)
1662 if (useGpuForcesHaloExchange)
1664 if (domainWork.haveCpuLocalForceWork)
1666 stateGpu->copyForcesToGpu(forceOut.forceWithShiftForces().force(), AtomLocality::Local);
1668 communicateGpuHaloForces(*cr, domainWork.haveCpuLocalForceWork);
1672 if (stepWork.useGpuFBufferOps)
1674 stateGpu->waitForcesReadyOnHost(AtomLocality::NonLocal);
1676 dd_move_f(cr->dd, &forceOut.forceWithShiftForces(), wcycle);
1681 // With both nonbonded and PME offloaded a GPU on the same rank, we use
1682 // an alternating wait/reduction scheme.
1683 bool alternateGpuWait = (!c_disableAlternatingWait && useGpuPmeOnThisRank && simulationWork.useGpuNonbonded
1684 && !DOMAINDECOMP(cr) && !stepWork.useGpuFBufferOps);
1685 if (alternateGpuWait)
1687 alternatePmeNbGpuWaitReduce(fr->nbv.get(), fr->pmedata, &forceOut, enerd, stepWork, wcycle);
1690 if (!alternateGpuWait && useGpuPmeOnThisRank)
1692 pme_gpu_wait_and_reduce(fr->pmedata, stepWork, wcycle, &forceOut.forceWithVirial(), enerd);
1695 /* Wait for local GPU NB outputs on the non-alternating wait path */
1696 if (!alternateGpuWait && simulationWork.useGpuNonbonded)
1698 /* Measured overhead on CUDA and OpenCL with(out) GPU sharing
1699 * is between 0.5 and 1.5 Mcycles. So 2 MCycles is an overestimate,
1700 * but even with a step of 0.1 ms the difference is less than 1%
1703 const float gpuWaitApiOverheadMargin = 2e6F; /* cycles */
1704 const float waitCycles = Nbnxm::gpu_wait_finish_task(
1705 nbv->gpu_nbv, stepWork, AtomLocality::Local, enerd->grpp.ener[egLJSR].data(),
1706 enerd->grpp.ener[egCOULSR].data(), forceOut.forceWithShiftForces().shiftForces(), wcycle);
1708 if (ddBalanceRegionHandler.useBalancingRegion())
1710 DdBalanceRegionWaitedForGpu waitedForGpu = DdBalanceRegionWaitedForGpu::yes;
1711 if (stepWork.computeForces && waitCycles <= gpuWaitApiOverheadMargin)
1713 /* We measured few cycles, it could be that the kernel
1714 * and transfer finished earlier and there was no actual
1715 * wait time, only API call overhead.
1716 * Then the actual time could be anywhere between 0 and
1717 * cycles_wait_est. We will use half of cycles_wait_est.
1719 waitedForGpu = DdBalanceRegionWaitedForGpu::no;
1721 ddBalanceRegionHandler.closeAfterForceComputationGpu(cycles_wait_gpu, waitedForGpu);
1725 if (fr->nbv->emulateGpu())
1727 // NOTE: emulation kernel is not included in the balancing region,
1728 // but emulation mode does not target performance anyway
1729 wallcycle_start_nocount(wcycle, ewcFORCE);
1730 do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::Local,
1731 DOMAINDECOMP(cr) ? enbvClearFNo : enbvClearFYes, step, nrnb, wcycle);
1732 wallcycle_stop(wcycle, ewcFORCE);
1735 // If on GPU PME-PP comms or GPU update path, receive forces from PME before GPU buffer ops
1736 // TODO refactor this and unify with below default-path call to the same function
1737 if (PAR(cr) && !thisRankHasDuty(cr, DUTY_PME)
1738 && (simulationWork.useGpuPmePpCommunication || simulationWork.useGpuUpdate))
1740 /* In case of node-splitting, the PP nodes receive the long-range
1741 * forces, virial and energy from the PME nodes here.
1743 pme_receive_force_ener(fr, cr, &forceOut.forceWithVirial(), enerd,
1744 simulationWork.useGpuPmePpCommunication,
1745 stepWork.useGpuPmeFReduction, wcycle);
1749 /* Do the nonbonded GPU (or emulation) force buffer reduction
1750 * on the non-alternating path. */
1751 if (useOrEmulateGpuNb && !alternateGpuWait)
1753 // TODO simplify the below conditionals. Pass buffer and sync pointers at init stage rather than here. Unify getter fns for sameGPU/otherGPU cases.
1755 stepWork.useGpuPmeFReduction
1756 ? (thisRankHasDuty(cr, DUTY_PME) ? pme_gpu_get_device_f(fr->pmedata)
1757 : // PME force buffer on same GPU
1758 fr->pmePpCommGpu->getGpuForceStagingPtr()) // buffer received from other GPU
1759 : nullptr; // PME reduction not active on GPU
1761 GpuEventSynchronizer* const pmeSynchronizer =
1762 stepWork.useGpuPmeFReduction
1763 ? (thisRankHasDuty(cr, DUTY_PME) ? pme_gpu_get_f_ready_synchronizer(fr->pmedata)
1764 : // PME force buffer on same GPU
1765 static_cast<GpuEventSynchronizer*>(
1766 fr->pmePpCommGpu->getForcesReadySynchronizer())) // buffer received from other GPU
1767 : nullptr; // PME reduction not active on GPU
1769 gmx::FixedCapacityVector<GpuEventSynchronizer*, 3> dependencyList;
1771 if (stepWork.useGpuPmeFReduction)
1773 dependencyList.push_back(pmeSynchronizer);
1776 gmx::ArrayRef<gmx::RVec> forceWithShift = forceOut.forceWithShiftForces().force();
1778 if (stepWork.useGpuFBufferOps)
1780 // Flag to specify whether the CPU force buffer has contributions to
1781 // local atoms. This depends on whether there are CPU-based force tasks
1782 // or when DD is active the halo exchange has resulted in contributions
1783 // from the non-local part.
1784 const bool haveLocalForceContribInCpuBuffer =
1785 (domainWork.haveCpuLocalForceWork || havePPDomainDecomposition(cr));
1787 // TODO: move these steps as early as possible:
1788 // - CPU f H2D should be as soon as all CPU-side forces are done
1789 // - wait for force reduction does not need to block host (at least not here, it's sufficient to wait
1790 // before the next CPU task that consumes the forces: vsite spread or update)
1791 // - copy is not perfomed if GPU force halo exchange is active, because it would overwrite the result
1792 // of the halo exchange. In that case the copy is instead performed above, before the exchange.
1793 // These should be unified.
1794 if (haveLocalForceContribInCpuBuffer && !useGpuForcesHaloExchange)
1796 // Note: AtomLocality::All is used for the non-DD case because, as in this
1797 // case copyForcesToGpu() uses a separate stream, it allows overlap of
1798 // CPU force H2D with GPU force tasks on all streams including those in the
1799 // local stream which would otherwise be implicit dependencies for the
1800 // transfer and would not overlap.
1801 auto locality = havePPDomainDecomposition(cr) ? AtomLocality::Local : AtomLocality::All;
1803 stateGpu->copyForcesToGpu(forceWithShift, locality);
1804 dependencyList.push_back(
1805 stateGpu->getForcesReadyOnDeviceEvent(locality, stepWork.useGpuFBufferOps));
1807 if (useGpuForcesHaloExchange)
1809 dependencyList.push_back(cr->dd->gpuHaloExchange[0]->getForcesReadyOnDeviceEvent());
1811 nbv->atomdata_add_nbat_f_to_f_gpu(AtomLocality::Local, stateGpu->getForces(), pmeForcePtr,
1812 dependencyList, stepWork.useGpuPmeFReduction,
1813 haveLocalForceContribInCpuBuffer);
1814 // Copy forces to host if they are needed for update or if virtual sites are enabled.
1815 // If there are vsites, we need to copy forces every step to spread vsite forces on host.
1816 // TODO: When the output flags will be included in step workload, this copy can be combined with the
1817 // copy call done in sim_utils(...) for the output.
1818 // NOTE: If there are virtual sites, the forces are modified on host after this D2H copy. Hence,
1819 // they should not be copied in do_md(...) for the output.
1820 if (!simulationWork.useGpuUpdate || vsite)
1822 stateGpu->copyForcesFromGpu(forceWithShift, AtomLocality::Local);
1823 stateGpu->waitForcesReadyOnHost(AtomLocality::Local);
1828 nbv->atomdata_add_nbat_f_to_f(AtomLocality::Local, forceWithShift);
1832 launchGpuEndOfStepTasks(nbv, fr->gpuBonded, fr->pmedata, enerd, *runScheduleWork,
1833 useGpuPmeOnThisRank, step, wcycle);
1835 if (DOMAINDECOMP(cr))
1837 dd_force_flop_stop(cr->dd, nrnb);
1840 if (stepWork.computeForces)
1842 postProcessForceWithShiftForces(nrnb, wcycle, box, x.unpaddedArrayRef(), &forceOut,
1843 vir_force, *mdatoms, *fr, vsite, stepWork);
1846 // TODO refactor this and unify with above GPU PME-PP / GPU update path call to the same function
1847 if (PAR(cr) && !thisRankHasDuty(cr, DUTY_PME) && !simulationWork.useGpuPmePpCommunication
1848 && !simulationWork.useGpuUpdate)
1850 /* In case of node-splitting, the PP nodes receive the long-range
1851 * forces, virial and energy from the PME nodes here.
1853 pme_receive_force_ener(fr, cr, &forceOut.forceWithVirial(), enerd,
1854 simulationWork.useGpuPmePpCommunication, false, wcycle);
1857 if (stepWork.computeForces)
1859 postProcessForces(cr, step, nrnb, wcycle, box, x.unpaddedArrayRef(), &forceOut, vir_force,
1860 mdatoms, fr, vsite, stepWork);
1863 if (stepWork.computeEnergy)
1865 /* Sum the potential energy terms from group contributions */
1866 sum_epot(&(enerd->grpp), enerd->term);
1868 if (!EI_TPI(inputrec->eI))
1870 checkPotentialEnergyValidity(step, *enerd, *inputrec);
1874 /* In case we don't have constraints and are using GPUs, the next balancing
1875 * region starts here.
1876 * Some "special" work at the end of do_force_cuts?, such as vsite spread,
1877 * virial calculation and COM pulling, is not thus not included in
1878 * the balance timing, which is ok as most tasks do communication.
1880 ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::no);